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

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 criteria, 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, and show how it can be addressed efficiently using two-party secure computation (2PC). To that end, we design simple and efficient 2PC protocols for addressing this gap, implement them, and perform end-to-end testing on standard datasets such as MNIST and CIFAR-10. Our empirical results indicate that the added security benefit essentially comes 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.

• Anti abuse