Research On Local Differential Privacy Preserving Methods For Spatio-temporal Data Aggregation

The rapid development of Intelligent Perception,Multi-level Communication Networks and Edge Computing technologies,as well as the widespread popularity of mobile devices with sensing and positioning capabilities provide strong technical support for collection,transmission and analysis of fine-grained spatio-temporal data.Location-based service providers continuously collect and analyze the fine-grained individuals' spatio-temporal data to improve service quality and provide users with personalized services.However,individuals' spatio-temporal data is closely related to daily life and can be easily used to infer work locations,private addresses,and behavioral activities,etc.Since the spatio-temporal data has the highly sensitive characteristics,it will bring more prominent privacy issues in the collection and aggregation of spatio-temporal data.How to meet the needs of service providers to obtain valuable information while protecting the privacy of personal spatiotemporal data will encounter a huge challenge.In recent years,local differential privacy has gradually become a standard privacy preservation technology for data aggregation.The traditional central differential privacy is usually applied to the privacy release of aggregated data.Distributed differential privacy is realized by combining the traditional differential privacy with cryptography technology to be adopted for private data aggregation.Similar to traditional cryptographic-based secure multi-party computing methods,distributed differential privacy will also bring into excessive resource overhead and complex system structure.Local differential privacy is a strict,provable and practical privacy preservation technology with low resource overhead.It does not need to assume that the data collector is trustable and does not rely on the participation of trusted third parties.This technology has been widely deployed by Google Inc.,Apple Inc.,and Microsoft Inc.There has been a batch of research works in locally differentially private data aggregation,but there is a lack of research on the instability of sample complexity(participating users)and spatiotemporal data with temporal and spatial correlation,so this thesis considers doing conduct research from the following three aspects: locally differentially private spatial location aggregation under the small sample,locally differentially private spatio-temporal data aggregation,and locally differentially private spatio-temporal infinite stream data.Data aggregation with local differential privacy is a method that relies heavily on the amount of sample data.Fewer participating users(e.g.small number of participating users recruited by Mobile Crowd Sensing)leads to insufficient sample size and reduces the utility of statistical distribution.In order to ensure the statistical validity of the local differential privacy distribution estimation,this thesis adopts the Bayesian multipledummies technology based on the k-Subset mechanism and iterative Bayes technology to develop the private spatial location data aggregation protocol(PLAP).However,PLAP under a large safe region(domain)will bring relatively large communication overhead and participating users have different privacy requirements in real scenarios.For this,this thesis adopts PLAP and a public spatial taxonomy to develop and implement the solution of communication efficiency-first private spatial location aggregation and the solution of privacy-first private spatial location aggregation that meet personalized privacy requirements.The experimental evaluation results show that the two spatial location aggregation methods proposed in this thesis have a great improvement in term of utility compared with the existing methods.In the research of locally differentially private data aggregation,the existing differential privacy preservating data aggregation analysis methods lack sufficient exploration for spatial-temporal data with temporal correlation and spatial similarity and comfront the risk of privacy leakage.The definitions of traditional local differential privacy cannot capture the temporal correlation of spatio-temporal data.In response to this problem,this thesis proposes a novel event-level local differential privacy definition ,(?,?)-LDP for spatio-temporal data.Based on this theoretical basis and combined with the temporal correlation of data,we propose a ,(?,?)-LDP privacy preserving spatio-temporal data local perturbed algorithm based on HMM and Generalized Random Response,and develop a spatio-temporal data privacy aggregation method.The proposed algorithm can resist correlation inferencing attacks to prevent privacy leakage.Theoretical and experimental results show that the proposed method can achieve better utility.Spatio-temporal data is constantly being generated all the time,and there is an urgent need to design a privacy data aggregation method oriented to the spatial-teporal infinite streams.Existing private data aggregation methods do not consider spatial and temporal correlation.To solve this problem,considering that Clustering has the ability to capture feature correlation and -event privacy has the advantage of capturing infinite stream data,this thesis proposes a definition of window-level clustering-based local differential privacy,namely ,(w,?)-CLDP,which is used to preserving privacy for private infinite streams in the local setting.Based on this theoretical basis,and combined with 1-bit random response mechanism and generalized random response mechanism,this thesis proposes a ,(w,?)-CLDP privacy preserving spatio-temporal infinite data local perturbed algorithm based on spatial clusteing.Based on the proposed algorithm,we develop LBD and LBA local perturbed algorithms by using two privacy budget allocation strategies similar to traditional BD and BA,and their corresponding private spatio-temporal infinite data aggregation methods.The experimental evaluation results show that the proposed LBD-based method is subject to large fluctuations in utility under different privacy paremeters,but overall it is slightly better than the Uniform-based method,the LBA-based method can achieve the optimal utility of aggregation result.