Research On Differential Privacy Protection Method In Regression Algorithm

In the past few decades,digital information collected by businesses,organizations,and government agencies have produced a large number of data sets,and the rate of such data collection has increased dramatically over the past few years.Typically,data collectors or owners will publish or further analyze this data.However,most of the collected data sets contain private sensitive information.Even if data collectors or owners can apply several simple anonymization techniques to handle this sensitive information,it is still likely that this personal information will be disclosed.Therefore,privacy protection has become an urgent issue that needs to be solved.In the method of protecting personal privacy data,differential privacy as a newly proposed privacy definition can still avoid privacy leakage under the attack of maximum background knowledge,and will not cause excessive distortion of data.As differential privacy protection technology can provide such strict privacy protection effects,more and more researchers have begun to recognize and research the technology.The method of combining differential privacy and regression analysis is an important research direction.However,the related work is relatively few at present,and there are still problems such as low precision and high sensitivity.To this end,this paper uses the inherent characteristics of regression analysis to solve the above problems from two different regression analysis algorithms.The main contributions are as follows:1.For the problem that the differential privacy protection algorithm based on linear regression analysis has low precision,this paper proposes an improved differential privacy linear regression method based on genetic algorithm.Specifically,the method adopts the evolutionary idea of the natural selection of genetic algorithm and introduces the exponential mechanism in the process of selecting the optimal linear regression model parameters,which makes the whole process of the algorithm satisfy the differential privacy protection.At the same time,the method uses the genetic algorithm to access the characteristics of the sensitive data set only during the selection step,reasonably allocate the privacy budget,and improve the usability of the results.2.For the problem that the differential privacy protection algorithm based on decision tree regression analysis is sensitive,this paper proposes a differential privacy protection algorithm DiffETs based on ExtraTrees Model.Specifically,in the process of building each decision tree,we use the Laplace mechanism and the exponential mechanism to ensure that differential privacy is met.The exponential mechanism is used to select the best splitting feature when selecting the internal nodes of the decision tree,and the Laplace mechanism is used to add noise on the leaf nodes.For the proposed algorithm,we apply it in decision tree regression and decision tree classification respectively,which show that the algorithm has good accuracy and low sensitivity.The two differential privacy protection regression analysis algorithms proposed in this paper not only analyze the privacy of the algorithm from the theoretical aspect,but also verify that the algorithm satisfies the ?-differential privacy,and also carries out experimental comparison and analysis on the data set published by UCI.The experimental results show that compared with the existing differential privacy protection regression analysis algorithms,the two differential protection regression analysis algorithms proposed in this paper can obtain better accuracy on the basis of guaranteeing privacy protection and has a high value of the practical application.