![[2602.15919] Generalized Leverage Score for Scalable Assessment of Privacy Vulnerability](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.15919] Generalized Leverage Score for Scalable Assessment of Privacy Vulnerability
Summary
The paper presents a method for assessing privacy vulnerability in machine learning models using a generalized leverage score, enabling efficient evaluation without retraining models.
Why It Matters
As privacy concerns grow in AI, this research offers a novel approach to evaluate the risk of membership inference attacks on individual data points. By providing a scalable and computationally efficient metric, it enhances the ability of developers and researchers to safeguard user data without the overhead of extensive model retraining.
Key Takeaways
- Introduces a generalized leverage score for assessing privacy vulnerability.
- Enables evaluation of membership inference attack risks without model retraining.
- Demonstrates a strong correlation between the leverage score and MIA success rates.
- Offers a computationally efficient method applicable to deep learning models.
- Enhances understanding of data-dependent sensitivity in privacy assessments.
Statistics > Machine Learning arXiv:2602.15919 (stat) [Submitted on 17 Feb 2026] Title:Generalized Leverage Score for Scalable Assessment of Privacy Vulnerability Authors:Valentin Dorseuil (DI-ENS), Jamal Atif (CMAP), Olivier Cappé (DI-ENS) View a PDF of the paper titled Generalized Leverage Score for Scalable Assessment of Privacy Vulnerability, by Valentin Dorseuil (DI-ENS) and 2 other authors View PDF Abstract:Can the privacy vulnerability of individual data points be assessed without retraining models or explicitly simulating attacks? We answer affirmatively by showing that exposure to membership inference attack (MIA) is fundamentally governed by a data point's influence on the learned model. We formalize this in the linear setting by establishing a theoretical correspondence between individual MIA risk and the leverage score, identifying it as a principled metric for vulnerability. This characterization explains how data-dependent sensitivity translates into exposure, without the computational burden of training shadow models. Building on this, we propose a computationally efficient generalization of the leverage score for deep learning. Empirical evaluations confirm a strong correlation between the proposed score and MIA success, validating this metric as a practical surrogate for individual privacy risk assessment. Subjects: Machine Learning (stat.ML); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) Cite as: arXiv:2602.15919 [stat.ML] (or arXiv:2602.159...