Paper 2025/211 Original: eprint.iacr.org/2025/211 ↗

Prior-Based Label Differential Privacy via Secure Two-Party Computation

Abstract

Differential privacy (DP) is a fundamental technique used in machine learning (ML) training for protecting the privacy of sensitive individual user data. In the past few years, a new approach for combining prior-based Local Differential Privacy (LDP) mechanisms with a relaxed DP criterion, known as Label DP, has shown great promise in increasing the utility of the final trained model without compromising on the DP privacy budget. In this work, we identify a crucial privacy gap in the current implementations of these prior-based LDP mechanisms, namely the leakage of sensitive priors. We address the challenge of implementing such LDP mechanisms without leaking any information about the priors while preserving the efficiency and accuracy of the current insecure implementations. To that end, we design simple and efficient secure two-party computation (2PC) protocols for addressing this challenge, implement them, and perform end-to-end testing on standard datasets such as MNIST, CIFAR-10. Our empirical results indicate that the added security benefit essentially comes almost for free in the sense that the gap between the current insecure implementations and our proposed secure version, in terms of run-time overhead and accuracy degradation, is minimal. E.g., for CIFAR-10, with strong DP privacy parameter, the additional runtime due to 2PC is $\approx 3.9\%$ over WAN with $0.4\%$ decrease in accuracy over an insecure (non-2PC) approach.