Sharing and Generating Privacy-Preserving Spatio-Temporal Data Using Real-World Knowledge

Privacy-preserving spatio-temporal data sharing is vital in many machine learning and analysis tasks, such as managing disease spread or tailoring public services to a population's travel patterns. Current methods for data release are insufficiently accurate to provide meaningful utility, and they carry a high risk of deanonymization or membership inference attacks. These limitations and public concern over privacy and data protection has limited the extent to which data is shared. This work presents approaches generating and publishing spatio-temporal data, such as geographic locations and trajectories, with differential privacy. In the first solution, differentially private spatial data is generated using kernel density estimation and a road network-aware approach. In the second solution, a local differentially private mechanism is developed by perturbing hierarchically-structured, overlapping n-grams of trajectory data. Both of the solutions incorporate publicly available information, such as the road network or categories of places of interests, to enhance the utility of the output data without negatively affecting privacy or efficiency. Experiments with real-world data demonstrate that the private data can perform as well as the non-private data in a range of practical data science tasks.