Paper 2022/202

Through the Looking-Glass: Benchmarking Secure Multi-Party Computation Comparisons for ReLU's

Abdelrahaman Aly and Kashif Nawaz and Eugenio Salazar and Victor Sucasas

Abstract

Comparisons are a basic component of Rectified Linear Unit functions (ReLU's), ever more present in Machine Learning and specifically in Neural Networks. Motivated by the increasing interest on privacy-preserving Artificial Intelligence, we explore the state of the art of Multi-Party Computation (MPC) protocols for privacy preserving comparisons. We then introduce constant round variations of these protocols, which are compatible with customary fixed point arithmetic over MPC and, geared towards realistic ReLU implementations. Furthermore, we provide novel constructions, inspired by commonly used comparison protocols equipped with state of the art elements. We translate these results into practice and provide an open source library, compatible with current MPC software tools, showcasing our contribution. We include a comprehensive benchmarking on various adversarial settings. Finally, we offer conclusions about the viability of our protocols, when adopted for privacy-preserving Machine Learning.

Note: Fixing typos, in affiliation, introduced in the last update