2000 character limit reached
Private PAC learning implies finite Littlestone dimension (1806.00949v3)
Published 4 Jun 2018 in cs.LG, cs.AI, cs.CR, math.LO, and stat.ML
Abstract: We show that every approximately differentially private learning algorithm (possibly improper) for a class $H$ with Littlestone dimension~$d$ requires $\Omega\bigl(\log*(d)\bigr)$ examples. As a corollary it follows that the class of thresholds over $\mathbb{N}$ can not be learned in a private manner; this resolves open question due to [Bun et al., 2015, Feldman and Xiao, 2015]. We leave as an open question whether every class with a finite Littlestone dimension can be learned by an approximately differentially private algorithm.