Towards Enhanced Privacy and Usability in Trajectory Synthesis

Trajectory data contains critical information about individual behavior habits and moving patterns, making it invaluable for applications such as traffic planning, urban management, business decision-making, etc. However, sharing this data raises significant privacy concerns, as it can expose sensitive details about individuals. This has created a pressing need for synthetic, privacy-preserving trajectory data in location-based applications. While differential privacy techniques generate synthetic trajectories by adding noise to probabilistic models and resampling, they often produce coarsegrained results that fail to reflect realistic movement paths, leading to poor usability. Recently, deep learning approaches learn the distribution of real trajectories but are limited by the scarcity of large, real-world trajectory datasets needed for training. To tackle these challenges, we present AnonyTraj, a new trajectory publication mechanism that strikes a balance between privacy-preserving and data usability. Furthermore, AnonyTraj employs a fully convolutional neural network (FCN) that leverages map features to identify plausible paths given starting points and movement directions, significantly enhancing the accuracy of subsequent trajectory point sampling. By refining initial trajectories generated by a differentially private adaptive Markov model as a privacy-preserving post-processing step, AnonyTraj overcomes the issue of coarse-grained results and produces synthetic trajectories that are both privacy-preserving and highly usable. Notably, AnonyTraj improves data utility by at least 33% compared to current state-of-the-art methods.