Learning with User-Level Privacy

TL;DR

This paper develops algorithms for user-level differential privacy in various learning tasks, showing how privacy costs decrease with more user data or users, and proving their optimality.

Contribution

It introduces new algorithms for user-level DP that improve privacy-utility trade-offs and provides theoretical bounds demonstrating their minimax optimality.

Findings

Privacy cost decreases as O(1/√m) with user samples.

Privacy cost decreases as O(1/n) with number of users.

Algorithms are proven to be minimax optimal for mean estimation and convex optimization.

Abstract

We propose and analyze algorithms to solve a range of learning tasks under user-level differential privacy constraints. Rather than guaranteeing only the privacy of individual samples, user-level DP protects a user's entire contribution ($m \ge 1$ samples), providing more stringent but more realistic protection against information leaks. We show that for high-dimensional mean estimation, empirical risk minimization with smooth losses, stochastic convex optimization, and learning hypothesis classes with finite metric entropy, the privacy cost decreases as $O(1/\sqrt{m})$ as users provide more samples. In contrast, when increasing the number of users $n$, the privacy cost decreases at a faster $O(1/n)$ rate. We complement these results with lower bounds showing the minimax optimality of our algorithms for mean estimation and stochastic convex optimization. Our algorithms rely on novel techniques for private mean estimation in arbitrary dimension with error scaling as the concentration radius $\tau$ of the distribution rather than the entire range.