Privacy-preserving machine learning based on secure four-party computations

Abstract

The paper is devoted to the analysis of privacy-preserving machine learning systems based on the concept of secure four-party computations. The advantages of the four-party computations over the two- and three-party computations are analyzed. The definitions are given that express the resistance of secure multi-party computation protocols against an active intruder. The Tetrad system is considered as an example of a secure four-party computations system with advanced functionality that most fully implements the properties of security against intruder actions. The concept of computations implemented in Tetrad based on the idea of mixed use of arithmetic as well as Boolean and garbled circuits in the function computations is analyzed. The multi-level architecture of the system is considered. The protocols related to various levels of architecture are analyzed in detail. For this purpose, an analysis of various forms of secret sharing is carried out. After that the basic protocols with shared secrets are considered. The systematization of protocols is completed with the analysis of high-level protocols used for the confidential implementation of model learning and inference procedures. This set of protocols allows performing high-level operations that are standard for machine learning systems, but with shared secrets. The paper concludes considering the advantages, disadvantages, features and limitations of four-party secure computations protocols for privacy-preserving machine learning.

Keywords

• privacy-preserving machine learning, secure multi-party computations, secret sharing scheme.