Private Algorithms for High-Dimensional Gaussians
Thursday, June 20, 2019 - 9:00am - 9:50am
We present novel differentially private algorithms for estimation and hypothesis testing of data from a high-dimensional Gaussian. In contrast to most previous differentially private algorithms for high-dimensional data, the performance of our algorithms nearly matches that of the optimal non-private algorithms for many parameter regimes. That is, privacy often comes for free. Our algorithms are based on novel private preconditioning methods for reducing the sensitivity of the relevant estimators and test statistics.