Sharper Utility Bounds for Differentially Private Models

In this paper, by introducing Generalized Bernstein condition, we propose the first high probability excess population risk bound for differentially private algorithms under the assumptions -Lipschitz, -smooth, and Polyak-{\L}ojasiewicz condition, based on gradient perturbation method. If we replace the properties -Lipschitz and -smooth by -H{\"o}lder smoothness (which can be used in non-smooth setting), the high probability bound comes to w.r.t , which cannot achieve when . To solve this problem, we propose a variant of gradient perturbation method, \textbf{max-Normalized Gradient Perturbation} (m-NGP). We further show that by normalization, the high probability excess population risk bound under assumptions -H{\"o}lder smooth and Polyak-{\L}ojasiewicz condition can achieve , which is the first high probability excess population risk bound w.r.t for differentially private algorithms under non-smooth conditions. Moreover, we evaluate the performance of the new proposed algorithm m-NGP, the experimental results show that m-NGP improves the performance of the differentially private model over real datasets. It demonstrates that m-NGP improves the utility bound and the accuracy of the DP model on real datasets simultaneously.
View on arXiv