Query Release via Online Learning

Years and Authors of Summarized Original Work

• 2010; Hardt, Rothblum

• 2012; Hardt, Rothblum, Servedio

• 2012; Thaler, Ullman, Vadhan

• 2013; Ullman

• 2014; Chandrasekaran, Thaler, Ullman, Wan

• 2014; Bun, Ullman, Vadhan

Problem Definition

Our goal is to design differentially private algorithms to answer statistical queries on a sensitive database. We model the database $D = (x_{1},\ldots,x_{n}) \in (\{0,1\}^{d})^{n}$ as a collection of n records – one per individual – each consisting of d binary attributes. A differentially private algorithm is a randomized algorithm whose output distribution does not depend “significantly” on any one record of the database. The formal definition is as follows:

Definition 1 ([8])

An algorithm $\mathcal{A}:(\{0,1\}^{d})^{n}\rightarrow$$\mathcal{R}$ is $(\varepsilon,\delta )$-differentially private if for every pair of databases $D,D^{{\prime}}\in (\{0,1\}^{d})^{n}$ that differ on at most one row and every $S \subseteq \mathcal{R}$,

$$\displaystyle{\mat...