English

DP-Adam: Correcting DP Bias in Adam's Second Moment Estimation

Machine Learning 2023-04-25 v1 Cryptography and Security

Abstract

We observe that the traditional use of DP with the Adam optimizer introduces a bias in the second moment estimation, due to the addition of independent noise in the gradient computation. This bias leads to a different scaling for low variance parameter updates, that is inconsistent with the behavior of non-private Adam, and Adam's sign descent interpretation. Empirically, correcting the bias introduced by DP noise significantly improves the optimization performance of DP-Adam.

Keywords

Cite

@article{arxiv.2304.11208,
  title  = {DP-Adam: Correcting DP Bias in Adam's Second Moment Estimation},
  author = {Qiaoyue Tang and Mathias Lécuyer},
  journal= {arXiv preprint arXiv:2304.11208},
  year   = {2023}
}

Comments

Published at ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models

R2 v1 2026-06-28T10:14:10.133Z