Jalaj Upadhyay, Rutgers University

Abstract

Continual counting is one of the fundamental problems studied in the literature of differential privacy with applications ranging from histogram estimation, graph analysis, matrix approximation, online optimization, and first provably private large-scale deployment of federated learning by Google. In this talk, I will give a brief historical overview of the problem and discuss some recent advances in differentially private continual observation that achieves fine-grained error bounds. We will also discuss some of the implications of these results in large-scale deployment and operator algebra. I will end the talk with some open problems. 

Based on joint work with Monika Henzinger and Sarvagya Upadhyay (https://arxiv.org/abs/2211.05006 and https://arxiv.org/abs/2307.08970).

See: https://theory.cs.rutgers.edu/theory_seminar