English

Differentially Private Online-to-Batch for Smooth Losses

Machine Learning 2022-10-14 v1 Cryptography and Security

Abstract

We develop a new reduction that converts any online convex optimization algorithm suffering O(T)O(\sqrt{T}) regret into an ϵ\epsilon-differentially private stochastic convex optimization algorithm with the optimal convergence rate O~(1/T+d/ϵT)\tilde O(1/\sqrt{T} + \sqrt{d}/\epsilon T) on smooth losses in linear time, forming a direct analogy to the classical non-private "online-to-batch" conversion. By applying our techniques to more advanced adaptive online algorithms, we produce adaptive differentially private counterparts whose convergence rates depend on apriori unknown variances or parameter norms.

Keywords

Cite

@article{arxiv.2210.06593,
  title  = {Differentially Private Online-to-Batch for Smooth Losses},
  author = {Qinzi Zhang and Hoang Tran and Ashok Cutkosky},
  journal= {arXiv preprint arXiv:2210.06593},
  year   = {2022}
}
R2 v1 2026-06-28T03:29:37.889Z