Secure multi-party computations for privacy-preserving machine learning

Abstract

The paper is devoted to the analysis of privacy-preserving machine learning (PPML) systems based on secure multi-party computations. It reviews PPML systems, analyses the goals and objectives of its application. A generalized model of PPML architecture is proposed, reflecting the main functional blocks of such systems. The formulation of the problem of secure multi-party computation is considered. The descriptions of cryptographic primitives and protocols used to implement multi-party secure computation protocols, including garbled circuits, secret sharing schemes, and homomorphic encryption are given. The current PPML systems based on two-, three-, and four-party secure computations are analyzed. The main attention is paid to algorithmic aspects of systems, methods and protocols of securing information. Systems secure against semi-honest and active adversaries are considered, both based on universal modules for secure multi-party computations, and specialized ones designed to ensure the privacy of specific machine learning technologies, such as convolutional neural networks. We consider examples of implemented prototypes of several PPML systems. Based on the results of the analysis, conclusions are formulated about the features of the future PPML systems.