Differentially Private Image Classification by Learning Priors from Random Processes

In privacy-preserving machine learning, differentially private stochastic gradient descent (DP-SGD) performs worse than SGD due to per-sample gradient clipping and noise addition. A recent focus in private learning research is improving the performance of DP-SGD on private data by incorporating pri...

Full description

Saved in:
Bibliographic Details
Main Authors: Xinyu Tang, Ashwinee Panda, Vikash Sehwag, Prateek Mittal
Format: Article
Language:English
Published: Labor Dynamics Institute 2025-03-01
Series:The Journal of Privacy and Confidentiality
Subjects:
Online Access:https://journalprivacyconfidentiality.org/index.php/jpc/article/view/910
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849731381927608320
author Xinyu Tang
Ashwinee Panda
Vikash Sehwag
Prateek Mittal
author_facet Xinyu Tang
Ashwinee Panda
Vikash Sehwag
Prateek Mittal
author_sort Xinyu Tang
collection DOAJ
description In privacy-preserving machine learning, differentially private stochastic gradient descent (DP-SGD) performs worse than SGD due to per-sample gradient clipping and noise addition. A recent focus in private learning research is improving the performance of DP-SGD on private data by incorporating priors that are learned on real-world public data. In this work, we explore how we can improve the privacy-utility tradeoff of DP-SGD by learning priors from images generated by random processes and transferring these priors to private data. We propose DP-RandP, a three-phase approach. We attain new state-of-the-art accuracy when training from scratch on CIFAR10, CIFAR100, MedMNIST and ImageNet for a range of privacy budgets $\epsilon \in [1, 8]$. In particular, we improve the previous best reported accuracy on CIFAR10 from $60.6 \%$ to $72.3 \%$ for $\epsilon=1$.
format Article
id doaj-art-4bbb39ea350043a483a33c04af07e314
institution DOAJ
issn 2575-8527
language English
publishDate 2025-03-01
publisher Labor Dynamics Institute
record_format Article
series The Journal of Privacy and Confidentiality
spelling doaj-art-4bbb39ea350043a483a33c04af07e3142025-08-20T03:08:35ZengLabor Dynamics InstituteThe Journal of Privacy and Confidentiality2575-85272025-03-0115110.29012/jpc.910Differentially Private Image Classification by Learning Priors from Random ProcessesXinyu Tang0Ashwinee Panda1Vikash Sehwag2Prateek Mittal3Princeton UniversityPrinceton UniversityPrinceton UniversityPrinceton University In privacy-preserving machine learning, differentially private stochastic gradient descent (DP-SGD) performs worse than SGD due to per-sample gradient clipping and noise addition. A recent focus in private learning research is improving the performance of DP-SGD on private data by incorporating priors that are learned on real-world public data. In this work, we explore how we can improve the privacy-utility tradeoff of DP-SGD by learning priors from images generated by random processes and transferring these priors to private data. We propose DP-RandP, a three-phase approach. We attain new state-of-the-art accuracy when training from scratch on CIFAR10, CIFAR100, MedMNIST and ImageNet for a range of privacy budgets $\epsilon \in [1, 8]$. In particular, we improve the previous best reported accuracy on CIFAR10 from $60.6 \%$ to $72.3 \%$ for $\epsilon=1$. https://journalprivacyconfidentiality.org/index.php/jpc/article/view/910differential privacyimage classification
spellingShingle Xinyu Tang
Ashwinee Panda
Vikash Sehwag
Prateek Mittal
Differentially Private Image Classification by Learning Priors from Random Processes
The Journal of Privacy and Confidentiality
differential privacy
image classification
title Differentially Private Image Classification by Learning Priors from Random Processes
title_full Differentially Private Image Classification by Learning Priors from Random Processes
title_fullStr Differentially Private Image Classification by Learning Priors from Random Processes
title_full_unstemmed Differentially Private Image Classification by Learning Priors from Random Processes
title_short Differentially Private Image Classification by Learning Priors from Random Processes
title_sort differentially private image classification by learning priors from random processes
topic differential privacy
image classification
url https://journalprivacyconfidentiality.org/index.php/jpc/article/view/910
work_keys_str_mv AT xinyutang differentiallyprivateimageclassificationbylearningpriorsfromrandomprocesses
AT ashwineepanda differentiallyprivateimageclassificationbylearningpriorsfromrandomprocesses
AT vikashsehwag differentiallyprivateimageclassificationbylearningpriorsfromrandomprocesses
AT prateekmittal differentiallyprivateimageclassificationbylearningpriorsfromrandomprocesses