NeurIPS 2020

Locally private non-asymptotic testing of discrete distributions is faster using interactive mechanisms

Meta Review

This paper makes strong technical contributions to the literature on differentially private hypothesis testing. In particular, the paper addresses the problem of goodness-of-fit under local differential privacy, and provides optimal tests and separation rates (under L1 and L2 distances). These results significantly improve over prior work. The paper also gives efficient algorithms and test procedures for both the non-interactive and sequentially interactive privacy mechanisms. All the reviewers agree that the results are strong and the paper is technically rich.