PLRV-O : Advancing Differentially Private Deep Learning via Privacy Loss Random Variable Optimization Conference: Proceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security; October 13–17, 2025; Taipei, Taiwan Proceedings of the 2025 ACM SIGSAC Conference on Computer and C

Qin Yang, Nicholas Stout, Meisam Mohammady, Han Wang, Ayesha Samreen, Christopher J. Quinn, Yan Yan, Ashish Kundu, Yuan Hong · 2025

Read PLRV-O : Advancing Differentially Private Deep Learning via Privacy Loss Random Variable Optimization Conference: Proceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security; October 13–17, 2025; Taipei, Taiwan Proceedings of the 2025 ACM SIGSAC Conference on Computer and C by Qin Yang, Nicholas Stout, Meisam Mohammady, Han Wang, Ayesha Samreen, Christopher J. Quinn, Yan Yan, Ashish Kundu, Yuan Hong (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-06264.txt.

Description

Private deep learning usually buys privacy with one noise knob, and Qin Yang and eight co-authors widen that control in a 2025 paper. Standard private training perturbs gradients with Gaussian noise set by a single scale, tying privacy loss and accuracy loss together. Their framework, PLRV-O, searches a broad family of noise distributions whose privacy loss can be characterized tightly yet tuned more independently.

The noise adapts to model size, training length, sampling, and clipping, in training and fine-tuning alike. Under strict privacy budgets, gains are large: a fine-tuned vision transformer on CIFAR-10 and a large language model on SST-2 both far exceed their Gaussian-noise counterparts at the same budget. The authors frame privacy loss itself as the object to optimize.

Catalog record

CREATORQin Yang, Nicholas Stout, Meisam Mohammady, Han Wang, Ayesha Samreen, Christopher J. Quinn, Yan Yan, Ashish Kundu, Yuan Hong
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE117 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-06264

Browse the full library · About the Tower of Babel