Application Research Of Homomorphic Encryption In Privacy-preserving Distributed Machine Learning

In the field of artificial intelligence,with the rapid increase of data and computation,the computing model of machine learning has been extended from centralized to distributed computing.In distributed machine learning environment,data security and privacy protection are important issues of concern to both data owners and computing service providers.Homomorphic encryption technology has become one of the main technologies to achieve secure distributed machine learning because that it has the properties of provable security and supporting ciphertext computing,which has aroused extensive research and application.In this paper,BGV homomorphic encryption scheme is applied to encrypt model parameters in machine learning.A distributed machine learning platform that can protect privacy is designed and developed,which is called joint machine learning platform under privacy protection.The main tasks completed include such aspects:(1)The design idea and algorithm flow of BGV scheme are studied.The hierarchical structure of HElib software library that is the implementation of BGV,the design of key classes and methods are analyzed.The model definition and algorithm idea of LDA and Linear Regression are studied,and the execution process of training tasks in machine learning library Shark is analyzed.(2)The structure of general distributed machine learning platform is studied,and the specific work of privacy protection in distributed machine learning environment is analyzed.The structure design and working mechanism of privacy-preserving distributed machine learning system are studied,and the design strategy of privacy protection in distributed machine learning system is analyzed.(3)The definition of Joint Machine Learning(JML)is proposed,and the design idea of privacy protection in JML platform is studied.The application scenario of JML is analyzed.The architecture model of JML under privacy protection and the updating strategy of model parameters are designed.(4)According to the JML architecture model,the overall architecture of JML platform and the key structure of central server are designed.According to the working principle of the platform model parameter updating strategy,the specific functions of platform data fusion,encoding & decoding and model parameter upload & download are designed.(5)The system implementations based on Linear Discriminant algorithm and Linear Regression algorithm are completed respectively,and the Iris plant data set and Somerville happiness index survey data set are used for experimental verification.Experiments show that under the premise of privacy protection,the Joint Machine Learning Platform can obtain the training results consistent with the single machine environment and the testing results superior to the single machine environment through enough iterations.The above work designs and implements the theoretical model and workflow of Privacy-preserving Joint Machine Learning.Experiments show that Joint Machine Learning achieves better performance than single machine environment in the algorithm of iteratively updating model parameters using optimizer.It has far-reaching potential in the field of distributed machine learning environment and Internet of Things.