Research On Privacy Protection In Machine Learning

Machine learning algorithms are an excellent class of data mining technologies and have been widely used in practice to create useful models based on mass data.Many machine learning algorithms create many valuable models based on a large number of data,and solve many practical problems,such as numerical prediction,medical diagnosis,image classification,biological feature recognition,etc.These large-scale data include lots of sensitive information.Therefore,how to use machine learning algorithms to obtain valuable models without leaking sensitive information needs to be solved urgently.Linear regression algorithm,logical regression algorithm,and neural network algorithm are the most representative three machine learning algorithms.They are progressive in structure and also the basis of many machine learning algorithms.In this paper,we extend the threshold secret sharing scheme,construct a perfect secure multi-party computing system,and combine the secure computing and three machine learning algorithms to construct a machine learning algorithm that can ensure the threshold privacy security in the case of multi data sources.The new privacy protection scheme endows machine learning algorithms with the characteristics of threshold privacy security.Assuming that there are n participants,and the threshold value is set to t,threshold privacy security means that the algorithm can ensure privacy security when multiple participants are monitored or controlled,and ensure the normal operation of the algorithm when multiple participants are offline or even destroyed.As far as we know from the literature,this paper proposes a threshold privacy protection machine learning algorithm for the first time in multi data source scenarios.At the same time,we also use C++ language to realize the threshold privacy-preserving linear regression algorithm of privacy protection,the threshold privacy-preserving logic regression algorithm,and the threshold privacy-preserving neural network algorithm of threshold privacy,and do experiments on different data sets.Experimental results show that our privacy-preserving scheme can achieve the same effect as the original machine learning algorithm on the premise of privacy security.Compared with the existing literature,our solution is more efficient.In addition,the characteristics of threshold security make the scheme more applicable and more universal.