A study on privacy protection techniques for spatiotemporal data

The School of ITEE is hosting the following progress review seminar:

A study on privacy protection techniques for spatiotemporal data

Speaker: Ms Fengmei Jin
Host: A/Prof. Hongzhi Yin

Abstract: With the ubiquitous influence of GPS-enabled devices and localization techniques, various types of spatiotemporal data are generated on a daily basis and become publicly available over the Internet, including GPS-sampled vehicle trajectories, location check-ins from social media, and so on. On the other hand, techniques like trajectory data mining aim to discover useful knowledge from spatiotemporal data so that provide better-personalized services, which inevitably magnify the unicity and sensitivity of spatiotemporal data as well as trigger some privacy concerns.

In this thesis, we focus on two types of issues: 1) the uniqueness of trajectories may be utilized by reidentification attacks which threaten trajectory data releasing; and 2) the sensitivity of semantic POIs brings the opportunity to attribute attack and so affects the flexible use of location-based services. To this end, we first studied the linkability of taxi trajectories with complex features and implemented a powerful reidentification attack via trajectory signatures. We then experimentally evaluated the existing protection models of various privacy principles against reidentification, which leads to valuable insights. With a such solid foundation, we developed our DP-based frequency randomization mechanisms which overcome the limitations from the trade-off between privacy and utility to the underestimated recovery risks specific to trajectory data. Regarding the second privacy concern, we emphasize semantic-level privacy protection to avoid the leakage of visit purposes or personal interests (e.g., visiting a hospital). To do so, we proposed semantic-aware indistinguishability models associated with non-trivial mechanism designs, which provide theoretically quantitative privacy guarantees. Apart from statistic-based metrics, we conduct an application-oriented evaluation by deploying our perturbation mechanisms in the task of privacy-preserving contact tracing, in which the semantics of check-in records are of great importance to individual privacy.

Our research has enhanced the privacy protection for spatiotemporal data in both theory and practice with a desirable trade-off among privacy, utility, and efficiency.

Biography: Ms Fengmei Jin is currently a PhD candidate in Data Science Discipline under the supervision of A/Prof. Wen Hua and Prof. Xiaofang Zhou. She received her Bachelor of Engineering from Sun Yat-Sen University in 2016 and Master of Engineering from Renmin University of China in 2019. Her research interests include efficient indexing structure, spatiotemporal data privacy and trajectory-user linking.