Evaluation of data security in Machine Learning L models

Speaker: Elizabeth Gassiat, University Paris- Saclay

Abstract: Membership inference attacks (MIA) can reveal whether a particular data point was part of the training dataset, potentially exposing sensitive information about individuals. In this talk,  I will explain the fundamental statistical limitations associated with MIAs on machine learning models at large. This allows to understand why overfitting learning procedures can lead to vulnerability to MIAs. I will also present consequences of our theoretical findings: discretizing data or quantizing machine learning models improves the security of the learning procedure. This is joint work with Eric Aubinais, Pablo Piantanida, and in part with Philippe Fromont.

Biography: Elisabeth Gassiat obtained her PhD in 1988 and until 2025, she has been Professor in Paris-Saclay University (Laboratoire de Mathématiques d'Orsay) where she is now emeritus. Her main research interests are statistical learning, nonparametric statistics, mixture and hidden Markov modeling, Bayesian inference, and coding theory. 

Chair: Prof. Deborah Sulem