DP-TrajGAN: A privacy-aware trajectory generation model with differential privacy

Abstract

Open Data Processing Services (ODPS) offers vast storage capacity and excellent efficiency, which collects and stores a lot of data. As an essential component of ODPS, location-based services (LBS) are widely used in many aspects. However, LBS generates tens of thousands trajectories, which have a significant likelihood of revealing personal information. In order to address this kind of privacy concerns, a novel model, namely Privacy-Aware Trajectory Generation Model with Differential Privacy (DP-TrajGAN), is proposed in this paper. Firstly, the long short-term memory network (LSTM) is improved and introduced into in the generative adversarial network (GAN) to learn the original distribution. Subsequently, the privacy-preserving of GAN is further enhanced using differential privacy (DP) while retaining the original features of the trajectories, which is called DP-TrajGAN. Furthermore, the privacy budget allocation in the DP is modeled using the Partially Observable Markov Decision Process (POMDP), which takes into account the trade-off between privacy and utility. The experimental findings demonstrate that, when compared to other models, DP-TrajGAN can provide higher quality trajectories with retained statistical features and effective privacy preservation.