Research On Local Differential Privacy Optimization Method Based On Data Disturbance

With the rapid development of artificial intelligence,cloud computing and other technologies,service providers are providing users with convenient and efficient services.At the same time,they are collecting a large number of valuable and sensitive privacy data from users.How to meet the needs of service providers for user privacy data collection while ensuring that user privacy is not disclosed is one of the great challenges facing the era of big data.In recent years,differential privacy technology has gradually developed into the standard definition of privacy protection.It does not need to make any assumptions about the background of potential attackers.Compared with traditional privacy protection technologies,such as k-anonymity,differential privacy technology has stronger privacy protection.According to the different methods of data collection and disturbance,the differential privacy technology can be divided into centralized differential privacy,distributed differential privacy and local differential privacy.Local differential privacy is more widely used because it disturbs data at the client,does not require a trusted third party and has less computing overhead.Therefore,once the concept of local difference privacy has been proposed,a large number of theoretical achievements and practical applications have been rapidly accumulated.However,as a new research field,the existing work is not enough to analyze its theoretical limitations.Moreover,with the trends of the complexity of application scenarios,the personalization of users’ privacy needs and the diversity of user data types in the era of big data,the existing schemes have great room for expansion in terms of statistical effectiveness,complex application scenarios and multi type data.In the framework of local differential privacy,this dissertation first studies how to solve the problem of privacy budget allocation based on the existing random response mechanism when facing high-dimensional heterogeneous attribute data;Then,according to the advantages and disadvantages of different random response mechanisms,a combined local differential privacy mechanism is proposed,which can solve the privacy protection problem of data with different dimensions and different privacy levels;Finally,considering the personalization of users’ privacy needs,a perturbation method is proposed to improve the current statistical validity.Specifically,this dissertation has completed the following work.Firstly,we proposed a local differential privacy optimization method under the condition of high dimensional heterogeneous data.local differential mechanism is applied in the high dimensional heterogeneous data privacy is the first big problem of difference privacy budget allocation problem,so this dissertation presents a high-dimensional heterogeneous data under the condition of optimal budget allocation,privacy and privacy by mathematical theorems budget allocation problem into a solution under the conditions of the solution of nonlinear equations.The proposed privacy budget allocation scheme has good generality and can be applied to the existing local differential privacy mechanism to realize the privacy publication of high-dimensional data.Secondly,we presented a combined local differential privacy optimization method for complex application scenarios.Aiming at the complex application scenarios with different attribute dimensions and different privacy levels in the same data set,this dissertation proposes a combined local differential privacy mechanism based on the advantages and disadvantages of binary random response and polytropic random response mechanisms under different dimensions and different privacy levels.This mechanism has strong applicability,and can deal with the local privacy publishing under the three conditions of same/heterogeneous,dimension and privacy level,which are applied separately or jointly.Finally,we proposed a local differential privacy optimization method to meet the personalized privacy requirements.Different users have different privacy attitudes towards the same data.To solve this problem,this dissertation proposes a localized differential privacy optimization method to meet users’ personalized privacy needs,and realizes the controllability of users’ privacy.From the perspective of data analysts,a weighted combination estimation method is proposed for joint histogram estimation under different privacy levels,and it is analyzed and evaluated respectively based on the two criteria of mean square error minimization and maximum error minimization.Finally,by theoretical analysis and experimental comparison,it is proved that this method can significantly reduce the estimation error.