Abstract

I'll discuss connections between robustness to adversarial corruption and differential privacy in high-dimensional statistical estimation problems; highlights include (1) a black-box reduction from privacy to robustness and (2) the first computationally-efficient algorithms for learning high-dimensional Gaussians privately with nearly-optimal sample complexity (and robustness).

Based on joint works with Kristian Georgiev, Gautam Kamath, Mahbod Majid, and Shyam Narayanan