Aggregating And Analyzing Data With Local Differential Privacy

The aggregation and analysis of data is the foundation of the information rev-olution,which accelerates the process of internet service generation and cloud-side decision-making.User privacy issues have evolved from "purple mud" to today's user data with the changes of the times.Data privacy issues make data encountered various difficulties in the aggregation-analysis-sharing-transaction process,resulting in users do not want to upload data,service providers dare not share data such as the "data island"phenomenon.Users are becoming more and more aware of privacy protection,and laws and regulations have gradually increased the requirements for privacy.There is an ur-gent need for service providers to transform the protection of user privacy from a burden to market competitiveness,to achieve a protective version of the business privacy.Lo-cal differential privacy technology is a new privacy standard,quantifying privacy and providing privacy protection to users through rigorous mathematical proof,without the need for trusted third parties,so it is more suitable for the actual situation.Users only need to process the original data locally,upload the perturbed version,the real data does not leave the local,compared to other methods such as cryptography has a better user experience.In this dissertation,we designed two enhancements to protect privacy for each of the two common data types.In the CPM advertising business scenario,which adopts the guaranteed quantity sale method,the advertising seller predicts the future available exposure inventory of each ad position under each orientation constraint based on the historical exposure data,so as to sell advertising resources reasonably.Historical exposure data type is multi-dimensional continuous value,for this type of data,and in order to achieve the protection of privacy data sharing release and maximize data availability to build inventory struc-ture,in work one,by designing constraint conditions and a probability density function graph that is beneficial to clustering,and setting parameter values for special points,the general parameter values of the new local differential privacy algorithm are ob-tained,and this method is used to process exposure log.Finally based on the perturbed data,combined with calculations such as clustering and sampling,we design a privacy-preserving inventory estimation framework.Rigorous mathematical proof shows the protection of RPM privacy,and experiments on simulation and real data sets show that RPM is superior to other local differential privacy algorithms under certain conditions.In the automatic desensitization system of image unstructured data with rich se-mantic content,because of the privacy characteristics of the data itself,only the user's feature operation can be uploaded and the cloud side decision can be updated,but the feature operation will also reveal the user's privacy to a certain extent.In the second work,by encoding the picture and feature operation as the key value pair data,using the frequency statistics and mean calculation of local difference privacy respectively to achieve the analysis of key and value,and finally through regression fitting to achieve the binding of key values,to achieve the effect of end-cloud synchronization.The method uses Lasso regression,focusing on the opinions of most users,ignoring smaller param-eters,and finally updating the decision through continuous iteration.Experiments on simulated data sets prove the feasibility of the method,and the whole method is low coupling and highly migratory.