SOK: homomorphic encryption in machine learning

Abstract

The field of machine learning (ML) has become ubiquitous, with new systems and models being implemented in a diverse range of domains resulting in the widespread use of software-based training and inference on third-party cloud platforms. There is growing recognition that outsourcing and hosting machine learning applications in the cloud introduces vulnerabilities in privacy and security. This paper systematizes findings on machine learning and homomorphic encryption, a privacy-preserving technology that is gaining popularity, focusing on the existing performance gap and other related works to improve its efficiency. The effect of using different hardware platforms has been surveyed. Moreover, the possibilities of combining it with other privacy-preserving technologies are discussed. Key insights resulting from works both in the ML and security communities are identified and the effectiveness of various approaches have been evaluated. The need for standardization and more detailed benchmarks has also been highlighted.