Differentially Private Clipped-SGD: High-Probability Convergence with Arbitrary Clipping Level

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Khah, Saleh Vatan, Chezhegov, Savelii, Farahmand, Shahrokh, Horváth, Samuel, Gorbunov, Eduard
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916975629828096
author Khah, Saleh Vatan
Chezhegov, Savelii
Farahmand, Shahrokh
Horváth, Samuel
Gorbunov, Eduard
author_facet Khah, Saleh Vatan
Chezhegov, Savelii
Farahmand, Shahrokh
Horváth, Samuel
Gorbunov, Eduard
contents Gradient clipping is a fundamental tool in Deep Learning, improving the high-probability convergence of stochastic first-order methods like SGD, AdaGrad, and Adam under heavy-tailed noise, which is common in training large language models. It is also a crucial component of Differential Privacy (DP) mechanisms. However, existing high-probability convergence analyses typically require the clipping threshold to increase with the number of optimization steps, which is incompatible with standard DP mechanisms like the Gaussian mechanism. In this work, we close this gap by providing the first high-probability convergence analysis for DP-Clipped-SGD with a fixed clipping level, applicable to both convex and non-convex smooth optimization under heavy-tailed noise, characterized by a bounded central $α$-th moment assumption, $α\in (1,2]$. Our results show that, with a fixed clipping level, the method converges to a neighborhood of the optimal solution with a faster rate than the existing ones. The neighborhood can be balanced against the noise introduced by DP, providing a refined trade-off between convergence speed and privacy guarantees.
format Preprint
id arxiv_https___arxiv_org_abs_2507_23512
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Differentially Private Clipped-SGD: High-Probability Convergence with Arbitrary Clipping Level
Khah, Saleh Vatan
Chezhegov, Savelii
Farahmand, Shahrokh
Horváth, Samuel
Gorbunov, Eduard
Machine Learning
Optimization and Control
Gradient clipping is a fundamental tool in Deep Learning, improving the high-probability convergence of stochastic first-order methods like SGD, AdaGrad, and Adam under heavy-tailed noise, which is common in training large language models. It is also a crucial component of Differential Privacy (DP) mechanisms. However, existing high-probability convergence analyses typically require the clipping threshold to increase with the number of optimization steps, which is incompatible with standard DP mechanisms like the Gaussian mechanism. In this work, we close this gap by providing the first high-probability convergence analysis for DP-Clipped-SGD with a fixed clipping level, applicable to both convex and non-convex smooth optimization under heavy-tailed noise, characterized by a bounded central $α$-th moment assumption, $α\in (1,2]$. Our results show that, with a fixed clipping level, the method converges to a neighborhood of the optimal solution with a faster rate than the existing ones. The neighborhood can be balanced against the noise introduced by DP, providing a refined trade-off between convergence speed and privacy guarantees.
title Differentially Private Clipped-SGD: High-Probability Convergence with Arbitrary Clipping Level
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2507.23512