Fully homomorphic encryption in PPMLAn review

Abstract

Fully homomorphic encryption (FHE) in privacy-preserving machine learning (PPML) is a current area of research value, aiming to achieve the protection of users private data by applying the concept of full homomorphic encryption to machine learning privacy preservation. The integration of the two involves extensive model modifications and performance issues. The current difficulties mainly focus on how to improve encryption efficiency through hardware or software, and how to apply homomorphic encryption to neural network models such as RNN that process sequence data. This paper introduces this complex research field, outlines two machine learning service models (MLaas and AIaas) that are concerned by the industry, summarizes the most advanced research technologies based on these two models in recent years, and discusses the technical difficulties and future research directions. As a difficult problem that has never been overcome in cryptography in recent decades, homomorphic technology has received extensive attention from experts and scholars and ushered in new opportunities in the current explosive development of machine learning.