56

An Iterative Algorithm for Differentially Private kk-PCA with Adaptive Noise

Main:10 Pages
4 Figures
Bibliography:5 Pages
Appendix:31 Pages
Abstract

Given nn i.i.d. random matrices AiRd×dA_i \in \mathbb{R}^{d \times d} that share a common expectation Σ\Sigma, the objective of Differentially Private Stochastic PCA is to identify a subspace of dimension kk that captures the largest variance directions of Σ\Sigma, while preserving differential privacy (DP) of each individual AiA_i. Existing methods either (i) require the sample size nn to scale super-linearly with dimension dd, even under Gaussian assumptions on the AiA_i, or (ii) introduce excessive noise for DP even when the intrinsic randomness within AiA_i is small. Liu et al. (2022a) addressed these issues for sub-Gaussian data but only for estimating the top eigenvector (k=1k=1) using their algorithm DP-PCA. We propose the first algorithm capable of estimating the top kk eigenvectors for arbitrary kdk \leq d, whilst overcoming both limitations above. For k=1k=1 our algorithm matches the utility guarantees of DP-PCA, achieving near-optimal statistical error even when n= ⁣O~(d)n = \tilde{\!O}(d). We further provide a lower bound for general k>1k > 1, matching our upper bound up to a factor of kk, and experimentally demonstrate the advantages of our algorithm over comparable baselines.

View on arXiv
Comments on this paper