Utility-Enhanced Spatio-Temporal Trajectory Synthesis With Differential Privacy

Vehicle trajectory data play a crucial role in traffic management, traffic prediction, and location-based services. However, directly publishing raw trajectory data poses significant risks to user privacy. Differential privacy addresses this issue by publishing synthetic datasets. However, existing approaches often overlook the temporal characteristics of trajectory data or assume uniform time intervals. In reality, temporal information is essential for maintaining the utility of trajectory data, and the time intervals in real-world trajectories are typically non-uniform. This discrepancy leads to the loss of substantial temporal information in synthetic datasets, which severely undermines their utility. In this paper, we propose a method forSpatio-temporalTrajectorySynthesis withDifferentialPrivacy (DP-STS). By normalizing time intervals in input datasets,DP-STSmitigates the negative impact of irregular time intervals on synthetic data utility. WithinDP-STS, we introduce a spatio-temporal movement probability model and a start-position/start-time association matrix, both of which capture spatio-temporal correlations. These components effectively extract and preserve the spatio-temporal correlations in the raw data, substantially improving the authenticity and validity of the synthetic trajectories. Extensive experiments on real-world and generated datasets demonstrate thatDP-STSoutperforms state-of-the-art methods in utility preservation. Specifically, the proposed approach achieves an average performance improvement of 47%.