Technology Research On Security Enhancement And Data Utility Optimization Of Differentially Private Mechanisms

Differential privacy is one of the cutting-edge technologies in the field of privacy protection.Its security and data utility are the foundation for its wide application,so they deserves further study.In this dissertation,the data security,parameter security,and data utility are investigated,and a method is proposed to enhance the security of differentially private mechanisms as well as a method to optimize the data utility of differentially private mechanisms.The contributions of this dissertation are as follows.(1)In terms of data security of differentially private mechanisms,it is found that the linear property of queries may lead to the disclosure of private information.The linear property allows a query to be split into the sum of two queries.Well-designed queries based on background knowledge can make differentially private mechanisms unable to accurately compute the consumption of the privacy budget,resulting in disclosing private information of data records.In this dissertation,this conclusion is verified by constructing membership inference attacks.(2)In terms of parameters security,it is found that differentially private mechanisms cannot maintain the confidentiality of their parameters such as privacy budget.Based on background knowledge,the noise added by the differentially private mechanisms can be calculated through sending queries with known answers to differentially private mech-anisms.Through the obtained noise,parameters of the noise distribution can be esti-mated by parameter estimation methods in statistics,which means that differentially pri-vate mechanisms' parameters such as privacy budget can be estimated.Thus,even if designers and users of differentially private mechanisms want to maintain the confiden-tiality of parameters such as privacy budget,they can not make it.In this dissertation,this conclusion is verified by constructing parameter recovery attack.And,the possible impact of parameter recovery attack is also shown in this dissertation.(3)In terms of security enhancement,a method is proposed to construct a differential privacy mechanism utilizing non-zero-mean noise.Non-zero-mean noise can ensure that the mean of multiple answers to the same query is no longer an unbiased estimator of the true answer to that query.That is,the estimation error does not go to zero when estimating the true answer to a query by the mean of multiple answers generated by the differentially private mechanism to the query.(4)In terms of data utility,differentially private mechanisms based on personalized sampling are proposed.By introducing personalized sampling into differentially private mechanisms,the data utility of differentially private mechanisms is improved without reducing the privacy guarantee.In this dissertation,personalized sampling laplace mech-anism together with personalized sampling and aggregation mechanism is proposed.The research on the security and data utility of the differentially private mechanisms in this dissertation is useful to design a more secure differentially private mechanism,reduce the impact on applications to achieve differential privacy,and promote the appli-cation of differential privacy in more business scenarios.