Private Stochastic Optimization with Large Worst-Case Lipschitz Parameter

Explore differentially private stochastic optimization in a 58-minute lecture from the USC Probability and Statistics Seminar. Delve into the challenges of loss functions with extremely large worst-case Lipschitz parameters due to outliers. Discover near-optimal excess risk bounds that overcome limitations of uniform Lipschitz assumptions, scaling with k-th moment bounds instead. Examine asymptotically optimal results for convex and strongly convex losses, as well as novel approaches for non-convex Proximal-PL functions. Learn about accelerated algorithms for smooth losses with tight excess risk in practical scenarios. Gain insights into addressing heavy-tailed data and outliers in private optimization, with applications to real-world machine learning problems.