Privacy Level Evaluation Of Differential Privacy For Time Series Based On Filtering Theory

With the rapid improvement of information technology,data are produced and stored at an amazing speed.As a significant form of data storage and publishing,time series is prevalent in various fields,such as process monitoring,financial predict and traffic dispatching.As time series could carry much useful information,and it is beneficial to our daily lives,researchers have focused on time series data mining.However,individuals' sensitive information may be revealed by the data mining results.Due to privacy considerations,an approach which could release personal data while ensuring the safety of the sensitive information is highly desired.Although the current differential privacy preserving methods on correlated time series have made some progress,they are proposed under the assumption that time series is correlated according to a certain model or rule.The key of the privacy preserving research is to build an attacker-centric system,however,the current differential privacy preserving methods on correlated time series are not designed by protecting against a specific attack model,and the privacy level of them cannot be measured.Moreover,the privacy level of them cannot be measured and compared.To solve the above problems,this paper did a lot of research about these problems.This paper mainly concentrates on three parts:Firstly,since the distribution of the Laplace noise designed by current methods is independent and different from that of the original correlated series,an optimal filter can be designed as a unified attack model.The proposed attack model can filter out the Laplace noise series from the perturbed time series,and therefore the attackers can obtain the original data with a higher possibility.Moreover,the privacy level changes of the current methods through the designed attack model are theoretically analyzed,and the proposed attack model can be used as a unified measurement for the current methods.Furthermore,we can conclude from the experimental results that,the correlation of the time series do have influence on the privacy level of the current methods,and the proposed attack model can evaluate the effective privacy level and data utility of these methods.