Differentially Private Trajectory Protection Based on Spatial and Temporal Correlation
Yun-Cheng WU,Hong CHEN,Su-Yun ZHAO,Wen-Juan LIANG,Yao WU,Cui-Ping LI,Xiao-Ying ZHANG
DOI: https://doi.org/10.11897/SP.J.1016.2018.00309
2018-01-01
Chinese Journal of Computers
Abstract:In recent years,location based services (LBS) is becoming one of the most important ways for information retrieval in our daily life,and it has broad application prospects and great value.However,people's locations or trajectory may be disclosed when they continuously use LBS to retrieve point of interests.This privacy disclosure problem not only restricts the development of LBS,but also reduces the quality of service the users obtained.Recently,trajectory privacy protection has attracted more and more attention,such as cloaking based technique,perturbation based technique,and so on.However,existing techniques seldom consider the geo-spatial and temporal correlation of the locations between several timestamps,which might degrade the location privacy of users.In this paper,aiming at dealing with the trajectory privacy problem,we explore a popular paradigm for providing privacy with strong theoretical guarantees,differential privacy,which has recently gained significant attention for its robustness to known attacks,and define a new privacy model based on differential privacy for trajectory protection.Specifically,we firstly propose an algorithm (CPL) to calculate the privacy level of each location on the map according to geo-spatial correlation.This algorithm transforms the topology of map into an undirected weighted graph.Based on the initial sensitive locations and the corresponding pre-defined privacy levels that provided by users,CPL algorithm iteratively allocates the privacy level of a location to its adjacent locations by the edge weights,and computes the aggregated privacy levels for all other locations that are not in the set of initial sensitive locations.Secondly,we present a privacy model,called γ-trajectory privacy,that combines the privacy level and differential privacy budget.Fundamentally,for any location in a trajectory,this privacy model requires that the multiplication of privacy level that computed from CPL algorithm and differential privacy budget of this location should equal to γ.In other words,the higher the privacy level is,the lower the differential privacy budget should be.Therefore,we can use this privacy model as a guideline to determine the differential privacy budget for every location.Thirdly,we figure out that if we release locations of a trajectory according to a differentially private location perturbation algorithm independently (which is widely used in existing work),malicious adversaries may also compromise the location privacy by the temporal correlation between perturbed locations.Thus,we propose a differentially private location release mechanism (DPLRM) that considers the temporal correlation to protect the trajectory privacy of users.Specifically,we model the temporal correlation between user's true locations by Markov chain transition matrix,and define the DPLRM as an optimization problem by minimizing an objective function based on the total distance between the true locations and possible released locations.We also give a mathematical deduction to calculate the constraints for this optimization problem.Finally,we conduct extensive experiments on two real world datasets,and show that it is computational efficient for CPL algorithm to compute the privacy levels,and the performance of DPLRM algorithm is close to an optimal approach and better than an existing mechanism.