Overview of Federal Learning and Privacy Protection

Abstract

In recent years, federated learning technology has developed rapidly and been widely used in the field of data processing. This paper makes a comprehensive discussion on the privacy protection methods in federated learning,and makes a detailed analysis of three basic methods:data encryption,data perturbation and trusted hardware-based,introduces the principle of each method,and objectively analyzes their performance in practical applications. This paper gives a visual example to compare the principle of secure multi-party computing, which is easy for readers to understand.Aiming at homorphic, this paper first introduces the background of this method and explains the algorith with block diagram,and then compares and analyzes the advantages and disadvantages of somewhat homorphic encryption and fully homorphic encryption. The specific denoising mechanisms are divided into three types:Laplace mechanism,Gauss mechanism and exponential mechanism.This paper summarizes these three mechanisms according to the types of data that need to be denoised.Finally,the paper systematically expounds the privacy protection methods based on trusted hardware ,lists two typical schemes,TrustZone and SGX,and analyzes how they work. In this paper,the hot spots and development prospects are prospected.