Dictionary · A Survey on Differentially Private Machine Learning [Review Article].

For two datasets D and D' that differ in at most one element, a randomized algorithm $M$ guarantees \emph{$(\epsilon, \delta)$-differential privacy} for any subset of the output $S$ if $M$ satisfies: \begin{equation} Pr[M(D) \in S] \leq exp(\epsilon)*Pr[M(D') \in S] + \delta \end{equation} Furthermore, when $\delta = 0$ an algorithm M is said to guarantee \emph{$\epsilon$-differential privacy}