报告题目：Gaussian Differential Privacy andthe Edgeworth Accountant        

报告时间： 2022-03-24 10:00-11:30

报告地点：腾讯会议（ID：943 727 127）        

报告摘要：

Privacy-preserving data analysishas been put on a firm mathematical foundation since the introduction of differentialprivacy (DP) in 2006. This privacy definition, however, has some well-knownweaknesses: notably, it does not tightly handle composition. In this talk, wepropose a relaxation of DP that we term "f-DP", which has a number ofappealing properties and avoids some of the difficulties associated with priorrelaxations. This relaxation allows for lossless reasoning about compositionand post-processing, and notably, a direct way to analyze privacy amplificationby subsampling. We define a canonical single-parameter family of definitionswithin our class that is termed "Gaussian Differential Privacy",based on hypothesis testing of two shifted normal distributions. We prove thatthis family is focal to f-DP by introducing a central limit theorem, whichshows that the privacy guarantees of any hypothesis-testing based definition ofprivacy converge to Gaussian differential privacy in the limit undercomposition. From a non-asymptotic standpoint, we introduce the EdgeworthAccountant, an analytical approach to compose $f$-DP guarantees of privatealgorithms. Finally, we demonstrate the use of the tools we develop by givingan improved analysis of the privacy guarantees of noisy stochastic gradientdescent.

报告人: Weijie Su（University ofPennsylvania）        

WeijieSu is an Assistant Professor in the Wharton Statistics and Data ScienceDepartment and, by courtesy, in the Department of Computer and InformationScience, at the University of Pennsylvania. He is a co-director of PennResearch in Machine Learning. Prior to joining Penn, he received his Ph.D. instatistics from Stanford University in 2016 and his bachelor’s degree inmathematics from Peking University in 2011. His research interests spanmultiple testing, privacy-preserving data analysis, optimization, high-dimensionalstatistics, and deep learning theory. He is a recipient of the StanfordTheodore W. Anderson Dissertation Award in Theoretical Statistics in 2016, anNSF CAREER Award in 2019, an Alfred Sloan Research Fellowship in 2020, and theSociety for Industrial and Applied Mathematics (SIAM) Early Career Prize inData Science in 2022.