Research On Time Series Data Privacy Protection Technology For Smart Grid

With the development of smart grid and the popularization of smart terminal devices,smart grid data shows explosive growth.Different from the traditional relational data,the data collected by smart grid are mainly time-series data,which have inherent rules and often show characteristics such as periodicity,seasonality and trend.Due to the continuous maturity of data mining technology,the mining and processing of time-series data has become the basis and core of many research fields and commercial applications.However,while exploring the intrinsic value of data,time-series data mining also causes great risks to the privacy and security of users.Therefore,while preserving the utility of data,effectively preventing the leakage of privacy information of users due to inference attacks against sensitive information has become an issue that cannot be ignored in the development of smart grid.In order to solve the above problems,this thesis focuses on the following two aspects:For the smart grid data collection stage,the data is vulnerable to attacks,from the perspective of data intrinsic organization model,the similarity matching-based temporal data privacy protection algorithm is proposed;For the smart grid data mining stage,the privacy information leakage problem may be triggered,and from the perspective of attack time series prediction model,this thesis proposes an adversarial attack method based on importance measure.This thesis aims to protect user privacy while preserving the utility of data,and the main research work is summarized as follows:1.For the security risks such as privacy attacks caused by the fine-grained data collected by smart grid,this thesis designs a privacy protection method for time-series data with inference resistant prediction.Firstly,this thesis analyzes the collected timeseries data,explores the characteristics of the time-series data,and uses a time-series data prediction model to make inferential predictions and then assesses the risk of user privacy leakage.Secondly,with the analysis of the time-series data,this thesis discovered that attackers can perform sensitive information inference attacks by performing historical data on users.Finally,in order to resist privacy inference,this thesis proposes a method based on time series similarity matching to find the original time-series data that have an important impact on inference results,and then perturb the identified data to reduce the ability of model inference prediction,so as to achieve the purpose of privacy protection.The method achieves security protection of data with a small perturbation cost.2.For the data mining phase,which is prone to problems such as data privacy leakage,this thesis designs an adversarial attack method for time series prediction model oriented to data privacy protection.Firstly,the gradient information is used to generate global perturbation-based adversarial samples.Secondly,in order to further reduce the difference between the adversarial sample and the original data,this thesis proposes an adversarial attack method based on importance measure,which measures the importance of moments in the adversarial sample and selects only the important moments in the adversarial sample to perturb the original time series.The adversarial sample generation algorithm proposed in this thesis makes the performance of the time-series data prediction model on time-series adversarial samples significantly reduced,thus suppressing the inference attacks on sensitive time-series data.This thesis utilizes household electricity dataset and solar power dataset to verify the algorithm proposed in this thesis.The experimental results show that the privacy protection method of smart grid time-series data designed in this thesis can not only resist inference attacks and achieve privacy protection for users,but also better preserve the utility of data to achieve data security management.