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

Abstract

The paper is devoted to the analysis of privacy-preserving machine learning systems based on the concept of secure three-party computations. After general information about the purposes of secure multi-party computations and privacy-preserving machine learning, an overview of existing privacy-preserving machine learning systems and perspectives for their development is offered. An analysis of the work of leading foreign research teams allows to identify several criteria essential for evaluating privacy-preserving machine learning systems based on multi-party secure computations. A comparative analysis of privacy-preserving machine learning systems is carried out according to a dedicated system of criteria. The further subject of consideration is only systems based on three-party secure computations. The main attention is paid to the algorithmic aspects of the organization of such systems, the methods and protocols of information security implemented in them. Systems secure to various types of adversary are considered, both based on universal modules of secure two-party computations, and specialized ones designed to ensure the privacy of specific machine learning methods, such as neural networks. Examples of prototypes of such systems are considered in detail. Based on the results of the analysis, conclusions are made about the prospects for developing privacy-preserving machine learning systems, and the tasks of future research are described.

Keywords

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