Differential Privacy Protection Models And Algorithms For Dynamic Data In Complex Environments

Through collecting and publishing massive dynamic data,the data center enables data sensing and data mining of large-scale systems at the terminal and the cloud,respectively.The dynamic data is real-time,continuous,open and massive,which leads to problems such as degraded data utility,restricted data length,untrusted aggregators,and high computational pressure when publishing dynamic data.Combining new techniques and concepts of differential privacy,this thesis investigates privacy-preserving algorithms with high efficiency and data utility for publication models of dynamic data in complex environments.The main research of this thesis is as follows.1.To address the problem of data utility degradation for publishing finite data streams in the trusted environment,the finite data streams release problem is modeled as a state space,and the privacy of finite data streams is ensured by the Laplace mechanism.In order to improve the data utility,filter-like differential privacy algorithms are derived by filtering techniques and the post-processing immunity of differential privacy.In order to avoid the problem of filter scattering in the recursive computation of filter-like differential privacy algorithms,a square root unscented Kalman filter-based differential privacy algorithm is proposed by using the square root of the covariance matrix to participate in the recursive update of the prediction process and the correction process.2.To address the problem of restricted data length for publishing infinite data streams in the trusted environment,using the majorization-minimization framework,a differentially private majorization-minimization algorithm is proposed to improve data utility.In order to improve the dynamic performance of the algorithm,using the recursive technique,a dynamic differentially private robust Kalman filtering algorithm is proposed.In order to ensure the privacy of infinite data streams,the concept of the discounted differential privacy is introduced to relax the privacy concept of past data by exploiting the user’s temporal preference for infinite data streams,and discounted differentially private robust filtering algorithms are proposed.3.To address the problem of untrusted aggregators for publishing dynamic data in the untrusted environment,the Cram′er’s decomposition theorem is used to inject Gaussian noise into each user’s data in a distributed manner to ensure the privacy of dynamic data.In order to ensure the scalability of the algorithm,a distributed Gaussian mechanism is designed using the data perturbation technique.In order to improve the utility of dynamic data,combined with the immunity of differential privacy to post-processing methods,a Gaussian perturbation-based moving average filtering scheme is proposed,and the privacy of the scheme is proved by the sequential composition.4.The centralized logistic regression model is decomposed into multiple sub-models by the alternating direction method of multipliers for the cloud computing pressure problem in the untrusted environment.In order to avoid direct interaction of dynamic data from the edge nodes with the cloud server and to relieve the computational pressure on the cloud server,each sub-model is trained in a distributed manner at the edge nodes and the model parameters obtained from the training are uploaded to the cloud.To achieve differential privacy of model parameters at edge nodes,by calibrating the noise level of model parameters with ?2sensitivity,a distributed logistic Gaussian perturbation algorithm is proposed.In summary,this thesis investigates the privacy protection of dynamic data in complex environments using new techniques and concepts of differential privacy.Several differential privacy protection models are given for different types of dynamic data and environments,and the corresponding privacy protection algorithms are proposed.The derivation procedure and privacy proof are given for the proposed algorithms.The algorithms are evaluated and compared by experimental simulations,and the simulation results show the effectiveness of the algorithms.