Research On Privacy Protection Based On Linear Regression Machine Learning Algorithm

Machine learning algorithm is the core of artificial intelligence.Locally Weighted Linear Regression algorithm is a classic algorithm widely used in machine learning,which can be used for housing price trend prediction,weather prediction,etc.In order to obtain better training effect of locally weighted linear regression algorithm,more training data are needed.Unfortunately,this data is often spread across multiple data owners.If the data is directly shared,it will lead to privacy leakage and other problems.In addition,in a complex communication environment,data owners will have problems such as unable to share data in real time and difficult interaction.This paper aims to solve the following problems:in a complex communication environment,multiple data owners can share data to improve the training effect of the locally weighted linear regression algorithm without compromising privacy.Therefore,this paper proposes three kinds of privacy-preserving schemes for locally weighted linear regression algorithm.At the same time,the feasibility of these schemes has been proved through experiments,and the computational efficiency of these three schemes has been gradually improved.The main work of this paper includes:1.The first scheme,a privacy-preserving locally weighted linear regression scheme based on stochastic gradient descent algorithm,is proposed.The scheme is based on the traditional cloud computing model,adopts the additive homomorphic encryption technology,and uses the classical machine learning model solution method-stochastic gradient descent method to solve the learning model.The experimental results show that the error between the model accuracy obtained by this scheme and that obtained by plaintext model accuracy can be as low as 10-4.2.The second scheme,a privacy-preserving locally weighted linear regression scheme based on Gaussian elimination and Jacobian iteration,is proposed.The scheme is also based on the traditional cloud computing model and adopts additive homomorphic encryption technology.However,compared with the first scheme,the proposed scheme improves the computational efficiency by packaging,encrypting and recalculating multiple plaintexts.Gaussian elimination method and Jacobian iteration method are used to solve the learning model,and the accuracy of the learning model is improved.3.The third scheme,a privacy-preseving locally weighted linear regression scheme based on federated learning model,is proposed.The scheme is based on federated learning model and adopts additive homomorphic encryption technology.Compared with the system models of the first two schemes,the federated learning model not only simplifies the model,but also further improves the computing efficiency.The experimental results show that the third scheme is much better than the first one and the second scheme in terms of the computational time cost of the whole learning model,and the application scenarios are more extensive.