Empirical Evaluation of Statistical Inference from Differentially-Private Contingency Tables

It is found that the theoretical guarantees associated with these differentially-private datasets do not always translate well into guarantees about the statistical inference on the synthetic datasets.

Abstract

In this paper, we evaluate empirically the quality of statistical inference from differentially-private synthetic contingency tables. We compare three methods: histogram perturbation, the Dirichlet-Multinomial synthesizer and the Hardt-Ligett-McSherry algorithm. We consider a goodness-of-fit test for models suitable to the real data, and a model selection procedure. We find that the theoretical guarantees associated with these differentially-private datasets do not always translate well into guarantees about the statistical inference on the synthetic datasets.