High-Precision Privacy-Preserving Real-Valued Function Evaluation

Abstract

We propose a novel multi-party computation protocol for evaluating continuous real-valued functions with high numerical precision. Our method is based on approximations with Fourier series and uses at most two rounds of communication during the online phase. For the offline phase, we propose a trusted-dealer and honest-but-curious aided solution, respectively. We apply our algorithm to train a logistic regression classifier via a variant of Newton’s method (known as IRLS) to compute unbalanced classification problems that detect rare events and cannot be solved using previously proposed privacy-preserving optimization algorithms (e.g., based on piecewise-linear approximations of the sigmoid function). Our protocol is efficient as it can be implemented using standard quadruple-precision floating point arithmetic. We report multiple experiments and provide a demo application that implements our algorithm for training a logistic regression model.