Spring 2019

Optimality in Locally Private Estimation and Learning

Tuesday, Apr. 9, 2019 3:00 pm3:45 pm PDT

Add to Calendar


John Duchi (Stanford University)

In this talk, I demonstrate fundamental lower bounds on locally private learning and estimation in all known models of differential privacy--differential, Renyi, approximate--for all values of privacy parameter and with all mechanisms of interaction. I will also discuss algorithms achieving these upper bounds, showing the first practical (and minimax optimal) methods for solving large-scale statistical learning and risk-minimization problems, with both theoretical and empirical evaluation.