While location trajectories represent a valuable data source for analyses and location-based services, they can reveal sensitive information, such as political and religious preferences [2]. For instance, researchers could determine which taxi drivers are practising Muslims based on the correlation between their routes and mandatory prayer times [3]. Moreover, De Montjoye et al. showed that only four spatio-temporal points suffice to identify 95% of individuals uniquely [4]. These examples illustrate that location trajectories require appropriate sanitation before being released for further analysis and that trajectory privacy represents a significant challenge in cybersecurity.

One special type of trajectories is semantic trajectories. Instead of location information, these trajectories consist of sequences of semantic locations, such as a credit card dataset containing the sequence of stores where a client used their card for payment. Similar privacy implications apply to semantic trajectories. For example, a shopping card history collected by a major supermarket might reveal whether a customer is pregnant, even before this knowledge is shared. This project considers the protection of such semantic trajectories.

Graph Neural Networks (GNNs) [1] are one of the most promising recent advancements in the field of Artificial Intelligence (AI), particularly for data represented as graphs. In contrast to standard location trajectories, semantic trajectories can be structured in such graph form due to the restricted domain. The application of GNNs in this trajectory privacy domain is relatively unexplored, presenting an opportunity to innovate in protecting privacy while maintaining the utility of semantic trajectory data. This work will focus on exploring GNN architectures and their suitability for encoding and transforming semantic trajectories in a privacy-preserving manner.

In this project, you will work on the fascinating intersection of Cybersecurity and AI. You will be among the first to explore the potential benefits of GNNs for trajectory privacy. The initial goal of this project is to design a GNN-based model for semantic trajectory anonymisation. One promising path could be using generative GNNs to generate similar yet distinct fake trajectories replacing sensitive data. However, this topic allows you to focus on the research direction you deem the most viable after your initial reading. So, you'll be able to own this project and shape it towards your own preferences under the guidance of your supervisors.

This project requires a solid understanding of deep learning, with practical experience with the PyTorch framework being highly beneficial. While prior knowledge of GNNs is an asset, it is not a prerequisite, as you will have the opportunity to develop this expertise during the project's initial phase. Familiarity with cybersecurity principles, especially in privacy protection, is advantageous but optional for this project.

References

  1. Z. Wu, et al., “A Comprehensive Survey on Graph Neural Networks,” IEEE Trans. Neural Netw. Learning Syst., vol. 32, no. 1, pp. 4–24, Jan. 2021, doi: 10.1109/TNNLS.2020.2978386.
  2. E. Buchholz, A. Abuadbba, S. Wang, S. Nepal, and S. S. Kanhere, “Reconstruction Attack on Differential Private Trajectory Protection Mechanisms,” in Proceedings of the 38th Annual Computer Security Applications Conference, in ACSAC ’22. New York, NY, USA: Association for Computing Machinery, December 2022, pp. 279–292. doi: 10.1145/3564625.3564628.
  3. L. Franceschi-Bicchierai, “Redditor cracks anonymous data trove to pinpoint Muslim cab drivers.” Accessed: Sep. 28, 2021. [Online]. 
  4. Y.-A. de Montjoye, et al., “Unique in the Crowd: The privacy bounds of human mobility,” Scientific Reports, vol. 3, no. 1, pp. 1–5, Dec. 2013, doi: 10.1038/srep01376. [Online].
School

Computer Science and Engineering

Research Areas

Privacy | Machine learning | Cyber security

The student will be part of a large group of security researchers working on topics that span the intersection of cyber-physical systems, machine learning, and cyber security. The group has sufficient resources available, such as high-performance servers. A senior PhD student, Erik, will provide support throughout the project.

The initial goal of this project is to design a GNN-based model for semantic trajectory anonymisation. One promising path could be using generative GNNs to generate similar yet distinct fake trajectories replacing sensitive data. However, this topic allows you to focus on the research direction you deem the most viable after your initial reading. So, you'll be able to own this project and shape it towards your own preferences under the guidance of your supervisors.

Default profile picture, avatar, photo placeholder. Vector illustration
Final year PhD student

  

opens in a new window

Recommended Reading

Background on Trajectory Privacy: "Reconstruction Attack on Differential Private Trajectory Protection Mechanisms.”
[Source code for above]

Graph Neural Networks in PyTorch:

Theoretical Knowledge of GNNs

Note: This reading is optional but serves as a guide to understanding the proposed topic.