Privacy-Preserving Mobile Crowdsourcing
Synopsis: An increasing number of applications use measurements collected from mobile devices and shared with a centralized entity, which raises legitimate privacy concerns. For example, end users may want to receive a location-based service but they may be concerned about sharing their exact location. The adversary can be the central entity itself or a third party that obtains access to all or part of the data stored at the central entity. The mobile user may apply privacy-preserving techniques to obfuscate the reported data, which, however, can cause a degradation of the service received.
The first goal of this proposal is to develop optimal obfuscation techniques, based on principled methodologies. In particular, we will consider three general frameworks for developing obfuscation techniques: (1) a Privacy-Utility optimization framework, primarily for cases where utility and privacy can be defined and tackled analytically; (2) Generative Adversarial Networks (GANs), primarily to produce data-driven obfuscation schemes using machine learning; and (3) Federated Learning, to enable distributed learning among multiple mobile devices and a central entity, while significantly reducing the information shared, from the training data to model parameters.
The second goal is to apply the framework to practical systems, including but not limited to: (1) spectrum sharing (2) signal maps and (3) leakage of personal information, all of which rely on crowdsourcing to collect network information and user data from mobile devices to a central entity. In each application scenario, we will formulate the application-specific problem and we will study and compare the privacy-utility tradeoff achieved by each of the three candidate obfuscation frameworks. Our approach will range from theory to practice: we will develop optimal obfuscation policies when possible, we will design practical heuristics, and we will evaluate them using real-world crowdsourced datasets as well as actual implementations on cellular testbeds and mobile devices
Current Personnel: Prof. Konstantinos Psounis USC, Dr. Mathew Clark, Aerospace Corp., Lillian Clark, USC, Jiang Zhang, USC, Pasit (Namo) Asavisanu, USC, Te Yi Kan, USC.
Past Personnel: Dr. Mathew Clark, Aerospace Corp., Lillian Clark, SpaceX.
Collaborative Institution: UCI.
Curent Collaborators: Prof. Athina Markopoulou, UCI.
Past Collaborators: Dr. Peter Kairouz, Google Brain.
Publications:
- Efficient Toxic Content Detection by Bootstrapping and Distilling Large Language Models, Jiang Zhang1, Qiong Wu, Yiming Xu, Cheng Cao, Zheng Du, and Konstantinos Psounis, in Proceedings of AAAI-24, 2024.
- Location Leakage in Federated Signal Maps. Evita Bakopoulou, Mengwei Yang, Jiang Zhang, Konstantinos Psounis, Athina Markopoulou, IEEE Transactions on Mobile Computing, October 2023, DOI: 10.1109/TMC.2023.3332034
- No Video Left Behind: A Utility-Preserving Obfuscation Approach for YouTube Recommendations, J. Zhang, H. Askari, K. Psounis, and Z. Shafiq, in Proceedings of PETS, 2023.
- A Unified Prediction Framework for Signal Maps: Not All Measurements are Created Equal. Alimpertis, Emmanouil; Markopoulou, Athina; Butts, Carter; Bakopoulou, Evita; Psounis, Konstantinos, IEEE Transactions on Mobile Computing, December 2022. DOI: 10.1109/TMC.2022.3221773
- HARPO: Learning to Subvert Online Behavioral Advertising, J. Zhang, K. Psounis, M. Haroon and Z. Shafiq, Network and Distributed Systems and Security Symposium (NDSS) 2022.
- Privacy-Utility Trades in Crowdsourced Mobile Network Data Obfuscation, J. Zhang, L. Clark, M. Clark, K. Psounis and P. Kairouz. Elsevier Computer Networks, October 2022.
- Optimizing Primary User Privacy in Spectrum Sharing Systems, M. Clark and K. Psounis. IEEE/ACM Transactions on Networking, April 2020.
- City-Wide Signal Strength Maps: Prediction with Random Forests, E. Alimpertis A. Markopoulou C. T. Butts and K. Psounis, in Proceedings of WWW, San Fransisco, CA, May 2019.
Broader Impact: This project advances the practices for handling crowdsourced mobile data, so as to provide privacy while still supporting the important services and economic activities enabled by crowdsourced mobile data. Specifically, the work on spectrum sharing privacy-performance trades provides optimal methods to spectrum share wireless bands among cellular providers and governmental agencies that protect the operation of the later from adversaries. Furthermore, the work on signal maps privacy-performance trades, provides a framework to compare different state of the art approaches, including Differential Privacy and Generative Adversarial Networks, to design efficient obfuscators that preserve the privacy of mobile data and users while allowing the creation of accurate signal maps.
Broadening Participation in Computing: Broadening participation in computing activities within USC include to grow Women in Cybersecurity (WiCys) Student Chapters at USC (and at the collaborative institution UCI) and to recruit minorities from LAUSD into USC’s CECS Undergraduate Program.