On the Performance of Differentially Private Optimization with Heavy-Tail Class Imbalance

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Tang, Qiaoyue, Zhiyanov, Alain, Lécuyer, Mathias
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908449223213056
author Tang, Qiaoyue
Zhiyanov, Alain
Lécuyer, Mathias
author_facet Tang, Qiaoyue
Zhiyanov, Alain
Lécuyer, Mathias
contents In this work, we analyze the optimization behaviour of common private learning optimization algorithms under heavy-tail class imbalanced distribution. We show that, in a stylized model, optimizing with Gradient Descent with differential privacy (DP-GD) suffers when learning low-frequency classes, whereas optimization algorithms that estimate second-order information do not. In particular, DP-AdamBC that removes the DP bias from estimating loss curvature is a crucial component to avoid the ill-condition caused by heavy-tail class imbalance, and empirically fits the data better with $\approx8\%$ and $\approx5\%$ increase in training accuracy when learning the least frequent classes on both controlled experiments and real data respectively.
format Preprint
id arxiv_https___arxiv_org_abs_2507_10536
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle On the Performance of Differentially Private Optimization with Heavy-Tail Class Imbalance
Tang, Qiaoyue
Zhiyanov, Alain
Lécuyer, Mathias
Machine Learning
In this work, we analyze the optimization behaviour of common private learning optimization algorithms under heavy-tail class imbalanced distribution. We show that, in a stylized model, optimizing with Gradient Descent with differential privacy (DP-GD) suffers when learning low-frequency classes, whereas optimization algorithms that estimate second-order information do not. In particular, DP-AdamBC that removes the DP bias from estimating loss curvature is a crucial component to avoid the ill-condition caused by heavy-tail class imbalance, and empirically fits the data better with $\approx8\%$ and $\approx5\%$ increase in training accuracy when learning the least frequent classes on both controlled experiments and real data respectively.
title On the Performance of Differentially Private Optimization with Heavy-Tail Class Imbalance
topic Machine Learning
url https://arxiv.org/abs/2507.10536