analiza aiqbench
This research addresses the accuracy gap in differentially private training by introducing NoiseCurve, a technique that leverages model curvature to optimize cross-iteration noise correlation. By allowing later gradient updates to better cancel out earlier privacy-induced noise, this approach significantly improves final model performance without compromising differential privacy guarantees. The method represents a substantive advancement in making high-quality deep learning models feasible under strict data protection constraints.
Correlating Cross-Iteration Noise for DP-SGD using Model Curvature
arXiv:2510.05416v3 Announce Type: replace Abstract: Differentially private stochastic gradient descent (DP-SGD) offers the promise of training deep learning models while mitigating many privacy risks. However, there is currently a large accuracy gap between DP-SGD and normal SGD training. This has resulted in different lines of research investigating orthogonal ways of improving privacy-preserving training. One such line of work, known as DP-MF, correlates the privacy noise across different iterations of stochastic gradient descent -- allowing later iterations to cancel out some of the noise added to earlier iterations. In this paper, we study how to improve this noise correlation. We propose a technique called NoiseCurve that uses model curvature, estimated from public unlabeled data, to improve the quality of this cross-iteration noise correlation. Our experiments on various datasets, models, and privacy parameters show that the noise correlations computed by NoiseCurve offer consistent and significant improvements in accuracy over the correlation scheme used by DP-MF.
Abstract:Differentially private stochastic gradient descent (DP-SGD) offers the promise of training deep learning models while mitigating many privacy risks. However, there is currently a large accuracy gap between DP-SGD and normal SGD training. This has resulted in different lines of research investigating orthogonal ways of improving privacy-preserving training. One such line of work, known as DP-MF, correlates the privacy noise across different iterations of stochastic gradient descent -- allowing later iterations to cancel out some of the noise added to earlier iterations. In this paper, we study how to improve this noise correlation. We propose a technique called NoiseCurve that uses model curvature, estimated from public unlabeled data, to improve the quality of this cross-iteration noise correlation. Our experiments on various datasets, models, and privacy parameters show that the noise correlations computed by NoiseCurve offer consistent and significant improvements in accuracy over the correlation scheme used by DP-MF.
Machine Learning (cs.LG)
arXiv:2510.05416 [cs.LG]
(or arXiv:2510.05416v3 [cs.LG] for this version)
Text integral sub licența deschisă a publicației · 159 cuvinte
Subiecte conexe
How can LLM RL Work Despite Information-Theoretic Inefficiency
LessWrong · 16 oct. 2026
Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
METR · 21 iul. 2026
Evaluating Red Team and Blue Team Capability for AI Control Research
LessWrong · 24 iul. 2026
Encoded Coordination on the Open Web
LessWrong · 23 sept. 2026