27
2

A Unified Characterization of Private Learnability via Graph Theory

Abstract

We provide a unified framework for characterizing pure and approximate differentially private (DP) learnability. The framework uses the language of graph theory: for a concept class H\mathcal{H}, we define the contradiction graph GG of H\mathcal{H}. Its vertices are realizable datasets, and two datasets S,SS,S' are connected by an edge if they contradict each other (i.e., there is a point xx that is labeled differently in SS and SS'). Our main finding is that the combinatorial structure of GG is deeply related to learning H\mathcal{H} under DP. Learning H\mathcal{H} under pure DP is captured by the fractional clique number of GG. Learning H\mathcal{H} under approximate DP is captured by the clique number of GG. Consequently, we identify graph-theoretic dimensions that characterize DP learnability: the clique dimension and fractional clique dimension. Along the way, we reveal properties of the contradiction graph which may be of independent interest. We also suggest several open questions and directions for future research.

View on arXiv
Comments on this paper