Research On Personalized Local Differential Privacy Scheme For Multidimensional Data

In the era of big data,companies and organizations have noticed the great value of multidimensional data and are collecting high-dimensional crowdsourced data to make datadriven decisions aggressively.However,these multidimensional data contain sensitive information of data owners.If the data owner shares his multidimensional data,it will cause privacy leakage.In recent years,local differential privacy has been found to have practical value in collecting and using data owners' data and protecting their privacy.In a local differential privacy scheme,the data owner perturbs his data before outsourcing it and sends its perturbed version to the server.In this way,the server cannot infer the actual data of the data owner accurately,but it can still estimate the overall distribution of the data accurately.However,these schemes often ignore data owners' personal privacy protection requirements.Considering the good characteristics of local differential privacy and the personalized privacy protection requirements of data owners,this paper studies the personalized local differential privacy scheme for multidimensional data:(1)Local differential privacy scheme supporting personalized privacy allocation for multidimensional data.In most existing local differential privacy schemes,data owners disturb their private data with the privacy budget set by the server.In fact,different data has different importance to each data owner,i.e.,each data owner has an unique privacy protection requirement for his data.If the privacy protection level provided by the server is lower than the data owner needs,the data owner may not be willing to share his data.Therefore,this paper considers data owners' personalized privacy protection requirements and proposes a multidimensional data local differential privacy scheme that supports personalized privacy allocation.In this scheme,we design a perturbation mechanism,called personalized multiple optimized unary encoding,to perturb data owners' data.Then,based on LASSO regression and information entropy,the joint distribution estimation algorithm of 6)-dimensional data is designed,which can achieve the similar result as the statistical analysis of perturbed data.(2)Social network graph generation scheme based on personalized local differential privacy for multidimensional data.Among the existing schemes of generating social network graphs based on local differential privacy,most of them only regard data owners' degree neighbor lists as the private data,and then use these perturbed degree neighbor lists to synthesize a social network graph.However,such a synthetic social network graph is not practical in real life because it is only a network graph composed of many nodes and edges.Even if some schemes take data owners' attribute data into account,the fixed privacy budgets of all degree data and attribute data preset by the server,which ignores data owner's personalized privacy protection requirements.To address it,this paper proposes a social network graph generation scheme based on personalized local differential privacy for multidimensional data.Firstly,a personalized randomized perturbation mechanism is designed to perturb data owners' degree data and attribute data.Then,an attribute joint distribution estimation algorithm based on expectation maximization is designed to estimate the data distribution.Finally,a seed graph creation mechanism and an optimization graph generation mechanism are designed to generate the social network graph according to the estimated data distribution.The social network graph generated by this scheme can not only provide personalized local differential privacy protection for data owners,but also generate efficient social network graph to serve the society.