Differentially Private Confidence Intervals for Empirical Risk Minimization

TL;DR

This paper develops methods for constructing confidence intervals in differentially private machine learning models, accounting for both sampling and privacy noise, compatible with existing privacy-preserving training mechanisms.

Contribution

It introduces algorithms for confidence intervals that satisfy differential privacy and concentrated differential privacy, applicable to models trained with objective and output perturbation.

Findings

Confidence intervals that satisfy differential privacy.

Compatibility with existing privacy-preserving mechanisms.

Applicability to various differentially private models.

Abstract

The process of data mining with differential privacy produces results that are affected by two types of noise: sampling noise due to data collection and privacy noise that is designed to prevent the reconstruction of sensitive information. In this paper, we consider the problem of designing confidence intervals for the parameters of a variety of differentially private machine learning models. The algorithms can provide confidence intervals that satisfy differential privacy (as well as the more recently proposed concentrated differential privacy) and can be used with existing differentially private mechanisms that train models using objective perturbation and output perturbation.