Asymptotically optimal and private statistical estimation

Abstract

Differential privacy is a definition of "privacy" for statistical databases. The definition is simple, yet it implies strong semantics even in the presence of an adversary with arbitrary auxiliary information about the database. In this talk, we discuss recent work on measuring the utility of differentially private analyses via the traditional yardsticks of statistical inference. Specifically, we discuss two differentially private estimators that, given i.i.d. samples from a probability distribution, converge to the correct answer at the same rate as the optimal nonprivate estimator. © 2009 Springer-Verlag.