Our project has resulted in the creation of secure and privacy-conscious dynamic spectrum sharing solutions for centralized spectrum management systems. We conducted a comprehensive study of privacy issues throughout the entire data analytics process within dynamic spectrum sharing systems. This investigation covered data collection, analytics, and utilization. Additionally, we jointly consider privacy issues with incentive design, resource allocation, data quality, and model robustness within the dynamic spectrum sharing framework.

One critical aspect of our work focused on protecting the location privacy of crowdsourced spectrum sensing agents involved in dynamic spectrum sharing. To accomplish this, we conducted a study to explore the intricate relationship between location privacy and the reconstruction of radio environment maps. Subsequently, we developed multiple methodologies aimed at preserving the location privacy of these agents during map reconstruction tasks.

In dynamic spectrum sensing, it's crucial to analyze data while safeguarding privacy. To improve data utility in privacy-preserving data analytics, we explored the concept of federated learning, which naturally integrates with data collection via mobile crowdsourcing. We investigated how to balance data privacy during training while preserving model accuracy. Our research covered two streams of federated learning: stochastic gradient descent (SGD) and alternating direction method of multipliers (ADMM). Specifically, for privacy-aware ADMM-based federated learning, we adapted classic ADMM approaches to the federated learning setting and proposed methods to enhance training accuracy with differential privacy guarantees. Theoretical analyses demonstrated the convergence rate and utility bounds of our proposed approaches across a wide range of learning objectives. Real-world experiments on diverse datasets illustrated the balance achieved between differential privacy and model utility. We also addressed practical issues in distributed learning, such as user diversity and the unique requirements of certain federated learning scenarios like model personalization.

Ensuring privacy is crucial for encouraging user participation in dynamic spectrum sensing systems. However, in many cases, some level of privacy trade-off is necessary to extract meaningful data analytical results. To address this challenge, we delved into incentive mechanism design. Specifically, we investigated how much the system should compensate users for contributing sensitive data and their valuable computing and communication resources. Employing game theory, we designed an effective incentive mechanism that selects users likely to provide reliable data and compensates them for the associated privacy risks. We formulated this as a two-stage Stackelberg game and demonstrated its effectiveness through extensive simulations. We also consider resource management issues, focusing on how to best utilize available resources to extract valuable information. We explored resource allocation to facilitate computing tasks at mobile edges, improving overall system efficiency.

Our project has wide-reaching societal and educational impacts. It addresses critical security and privacy challenges in aggressive spectrum sharing, improving spectrum efficiency and enabling innovative applications in various sectors. We've also contributed to education by training PhD and undergraduate students, including underrepresented minorities, and developing new course materials.

Furthermore, our research outcomes have led to a significant body of work, including 8 conference papers and 6 journal papers published in prestigious venues such as IEEE ICC, IEEE GLOBECOM, IEEE Internet of Things Journal, IEEE Transactions on Information Forensics and Security, IJCAI, and IEEE Open Journal of the Computer Society.

Last Modified: 10/14/2023
Modified by: Yanmin Gong