English

Sharper Utility Bounds for Differentially Private Models

Machine Learning 2022-10-19 v1

Abstract

In this paper, by introducing Generalized Bernstein condition, we propose the first O(pnϵ)\mathcal{O}\big(\frac{\sqrt{p}}{n\epsilon}\big) high probability excess population risk bound for differentially private algorithms under the assumptions GG-Lipschitz, LL-smooth, and Polyak-{\L}ojasiewicz condition, based on gradient perturbation method. If we replace the properties GG-Lipschitz and LL-smooth by α\alpha-H{\"o}lder smoothness (which can be used in non-smooth setting), the high probability bound comes to O(nα1+2α)\mathcal{O}\big(n^{-\frac{\alpha}{1+2\alpha}}\big) w.r.t nn, which cannot achieve O(1/n)\mathcal{O}\left(1/n\right) when α(0,1]\alpha\in(0,1]. To solve this problem, we propose a variant of gradient perturbation method, \textbf{max{1,g}\{1,g\}-Normalized Gradient Perturbation} (m-NGP). We further show that by normalization, the high probability excess population risk bound under assumptions α\alpha-H{\"o}lder smooth and Polyak-{\L}ojasiewicz condition can achieve O(pnϵ)\mathcal{O}\big(\frac{\sqrt{p}}{n\epsilon}\big), which is the first O(1/n)\mathcal{O}\left(1/n\right) high probability excess population risk bound w.r.t nn 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.

Cite

@article{arxiv.2204.10536,
  title  = {Sharper Utility Bounds for Differentially Private Models},
  author = {Yilin Kang and Yong Liu and Jian Li and Weiping Wang},
  journal= {arXiv preprint arXiv:2204.10536},
  year   = {2022}
}
R2 v1 2026-06-24T10:55:35.657Z