A New Linear Scaling Rule for Private Adaptive Hyperparameter Optimization

TL;DR

This paper introduces an adaptive hyperparameter optimization method for differentially private deep learning that efficiently estimates and scales hyperparameters, achieving state-of-the-art results across diverse tasks while preserving privacy.

Contribution

It presents a novel adaptive HPO technique that reduces privacy cost and runtime, enabling effective hyperparameter tuning under differential privacy constraints.

Findings

Achieved state-of-the-art performance on 22 benchmark tasks.

Effectively balances privacy cost and utility in hyperparameter tuning.

Applicable across various architectures and privacy budgets.

Abstract

An open problem in differentially private deep learning is hyperparameter optimization (HPO). DP-SGD introduces new hyperparameters and complicates existing ones, forcing researchers to painstakingly tune hyperparameters with hundreds of trials, which in turn makes it impossible to account for the privacy cost of HPO without destroying the utility. We propose an adaptive HPO method that uses cheap trials (in terms of privacy cost and runtime) to estimate optimal hyperparameters and scales them up. We obtain state-of-the-art performance on 22 benchmark tasks, across computer vision and natural language processing, across pretraining and finetuning, across architectures and a wide range of $\varepsilon \in [0.01,8.0]$, all while accounting for the privacy cost of HPO.