An Iterative Algorithm for Differentially Private -PCA with Adaptive Noise

Given i.i.d. random matrices that share a common expectation , the objective of Differentially Private Stochastic PCA is to identify a subspace of dimension that captures the largest variance directions of , while preserving differential privacy (DP) of each individual . Existing methods either (i) require the sample size to scale super-linearly with dimension , even under Gaussian assumptions on the , or (ii) introduce excessive noise for DP even when the intrinsic randomness within is small. Liu et al. (2022a) addressed these issues for sub-Gaussian data but only for estimating the top eigenvector () using their algorithm DP-PCA. We propose the first algorithm capable of estimating the top eigenvectors for arbitrary , whilst overcoming both limitations above. For our algorithm matches the utility guarantees of DP-PCA, achieving near-optimal statistical error even when . We further provide a lower bound for general , matching our upper bound up to a factor of , and experimentally demonstrate the advantages of our algorithm over comparable baselines.
View on arXiv