Differential privacy . . .-英文文献.pdf

Differential Privacy and Robust Statistics Cynthia Dwork Jing Lei∗ Microsoft Research Department of Statistics dwork@ University of California, Berkeley jinglei@ ABSTRACT In this work we turn to a field in which considerable care We show by means of several examples that robust statis- has already been given to ensuring some kind of probabilis- tical estimators present an excellent starting point for dif- tic insensitivity: robust statistics. We focus on parameter ferentially private estimators. Our algorithms use a new estimation, a branch of statistics that assumes data come paradigm for differentially private mechanisms, which we from a parameterized family of probability distributions and call Propose-Test-Release (PTR), and for which we give a makes inferences about the parameters of the distribution. formal definition and general composition theorems. Most well-known elementary statistical methods are para- metric. For example, given data samples x1 , . . . , xn drawn from a normal distribution F ∈ F = {N (θ, 1)}θ∈R , the goal Categories and Subject Descriptors is to estimate θ and a simple method averages the samples. H.2.0 [Information Systems]: Database Ma