Events

CDS Seminar: Finding (and Breaking) Privacy in Overparameterized Models

Speaker: Simone Bombari (NYU)

Location: 60 Fifth Avenue, Room 7th Floor Open Space

Date: Friday, October 16, 2026

Abstract: Increasing model size has been one of the secret sauces behind the remarkable progress of deep learning over the past decade. However, evidence suggests that larger models are also more inclined to memorize their training data, with clear implications for user privacy and copyright infringement.

In this talk, I will focus on a mathematically tractable family of overparameterized random feature models. I will first present evidence for a law of data reconstruction: full training-set reconstruction becomes possible once the model is sufficiently large. I will then discuss how reconstruction can be provably prevented through differentially private training, and show how the success of this method does not deteriorate with model size, challenging the opposing view in the field.

Bio: Simone Bombari is a CDS Faculty Fellow at NYU and a CCM Research Fellow at the Flatiron Institute, interested in the theoretical foundations of deep learning and trustworthy machine learning. He received his PhD from ISTA, advised by Marco Mondelli. During his PhD, he was an intern at AWS and G-Research, and he was supported by a Google PhD Fellowship.