DATA-CENTRIC AI FOR INTERACTION SECURITY AND PRIVACY IN THE INTERNET-OF-THINGS
In the realm of the Internet of Things (IoT), users, devices, and environments communicate and interact with each other, creating a web of complex interactions. This interconnected web of interactions makes the IoT a powerful tool for enhancing human experiences. However, it simultaneously presents substantial challenges in ensuring security and privacy amid interactions among users, devices, and environments. This dissertation investigates potential IoT interaction security and privacy issues by customizing data-centric AI algorithms. First, this dissertation studies complex interactions in smart homes where many interconnected smart devices are deployed. A graph learning-based threat detection system is designed to discover potential interactive threats across multiple smart home platforms. Second, considering smart home data privacy and data heterogeneity issues, a dynamic clustering-based federated graph learning framework is proposed to collaboratively train a threat detection model. Meanwhile, a Monte Carlo beam search-based method is designed to identify the interactive threat causes. Third, we explore the privacy issues behind the interactions between users and smartphones. Specifically, a potential bio-information leakage attack channel has been identified that utilizes near-ultrasound signals from a smartphone to recognize facial expressions based on a contrastive attention learning model. Fourth, we reveal two critical overprivileged issues in mobile activity sensing data generated from interactions between users and mobile devices: metadata-level and feature-level overprivileged issues. Correspondingly, we design the multi-grained data generation model to reconstruct mobile activity sensing data, so as to mitigate the privacy concerns behind the mobile sensing overprivileged issues. We have implemented and extensively evaluated the proposed threat detection model, federated model training method, acoustic-based expression recognition model, and privacy-preserving data reconstruction model in practical settings. This dissertation concludes with a discussion of future work. We highlight the potential challenges and opportunities associated with the applied AI techniques for addressing security and privacy issues in the IoT. This dissertation points out the pathway for future research in enhancing security and privacy to safeguard the interactions among users, devices, AI, and environments.
Read
- In Collections
-
Electronic Theses & Dissertations
- Copyright Status
- In Copyright
- Material Type
-
Theses
- Authors
-
Wang, Guangjing
- Thesis Advisors
-
Yan, Qiben
- Committee Members
-
Ross, Arun
O'Shea, Brian
Zhang, Mi
Zeng, Huacheng
- Date Published
-
2024
- Subjects
-
Computer science
- Program of Study
-
Computer Science - Doctor of Philosophy
- Degree Level
-
Doctoral
- Language
-
English
- Pages
- 176 pages
- Permalink
- https://doi.org/doi:10.25335/a6qd-k369