Beyond Worst Case Sensitivity in Private Data Analysis

Years and Authors of Summarized Original Work

• 2006; Cynthia Dwork, Frank McSherry, Kobbi Nissim, Adam Smith

• 2007; Kobbi Nissim, Sofya Raskhodnikova, Adam Smith

• 2009; Cynthia Dwork, Jing Lei

• 2013; Adam Smith, Abhradeep Thakurta

Problem Definition

Over the last few years, differential privacy [5, 6] has emerged as one of the most accepted notions of statistical data privacy. At a high level differential privacy ensures that from the output of an algorithm executed on a data set of potentially sensitive records, an adversary learns “almost” the same thing about an individual irrespective of his presence or absence in the data set. Formally, differential privacy is defined below (Definition 1). Setting the privacy parameters ε < 1 and $\delta \ll \frac{1} {n^{2}}$ ensures semantically meaningful privacy guarantees. For a detailed survey on the semantics of differential privacy, see [2, 3, 8, 9].

Definition 1 ((ε, δ)-differential privacy [5, 6])

We call two data sets D and D′ (with n...