Paper 2018/139

Faster Multiplication Triplet Generation from Homomorphic Encryption for Practical Privacy-Preserving Machine Learning under a Narrow Bandwidth

Wen-jie Lu and Jun Sakuma

Abstract

Machine learning algorithms are used by more and more online applications to improve the services. Machine learning-based online services are usually accessed by thousands of clients concurrently through a relatively narrow bandwidth, such as a WiFi network or a cell phone network. When applying secure computations to such online services, however, current methods for generating multiplication triplets might take a long time, especially when only a narrow bandwidth is available or large-scale matrices are involved in the computation. In this paper, we present a more practical method for generating multiplication triplets that are specified for additively shared matrices from homomorphic encryption. With our algorithmic and implement optimizations, our protocol is faster than and consumes less communication traffic than the existing methods. Experimental results show that, under a 100~Mbps network, our protocol took about $18.0$ seconds to generate triplets for matrices with more than $2.6\times 10^5$ entries. It was about $20 - 108$ times faster than existing methods. As the concrete example, we applied our protocol to two existing secure computation frameworks of machine learning, i.e., SecureML (S\&P'17) and MiniONN (CCS'17). Experimental results show that our method reduced about $74\% - 97\%$ of the triplet generation time of these frameworks when a narrow bandwidth was used.