Excellence In Research: Secure, Private, And Delay-Tolerant Non-Cooperative Games For Agent-Based Learning

This NSF EIR project aims to address security, privacy, and communication challenges in multi-agent systems by utilizing the framework of multi-cluster non-cooperative games. The project will bring a transformative change in autonomous and intelligent transportation systems. This is achieved by designing innovative solutions that bridge theory and real-world applications. The intellectual merit of the project includes designing a trust-based resiliency framework to mitigate Byzantine attacks even when Byzantine agents outnumber honest agents in the system, ensuring privacy without accuracy trade-offs or loss of privacy as time grows, and overcoming communication delays among agents. The broader impacts of the project include enhancing the undergraduate and graduate computer science curricula; providing hands-on experience for students in robotics and cybersecurity; training faculty members and students through workshops, seminars, and outreach engagements; and workforce development to meet the growing demand for skilled professionals in manufacturing, transportation, and aviation. Non-cooperative game theory provides a powerful framework for multi-cluster systems where agents cooperate within clusters and compete across them to achieve a generalized Nash equilibrium. However, several key challenges limit its deployment in the real-world dynamic environment. Existing approaches generally lack resilience to Byzantine attacks, often relying on the unrealistic assumption that honest agents outnumber malicious ones. Privacy preservation is also insufficiently addressed, as current methods frequently accumulate privacy loss over time, weakening long-term guarantees. Furthermore, most studies assume fixed communication delays, leaving arbitrary and time-varying delays largely unexplored. This project addresses these limitations by developing secure, privacy-preserving, and delay-resilient game-theoretic frameworks for clustered multi-agent systems. The research consists of three integrated thrusts. Thrust 1 leverages cyber-physical agent characteristics to detect and mitigate Byzantine attacks, even when adversarial agents form the majority. Thrust 2 advances privacy-preserving online learning through local differential privacy, using history-based average gradients, adaptive mirror-descent step sizes, and dynamic noise calibration to maintain strong privacy while preserving accuracy and efficiency. Thrust 3 develops delay-resilient online learning algorithms capable of handling both fixed and adversarial communication delays through optimistic hinting and delay-tolerant learning techniques. Collectively, these efforts establish a comprehensive framework for secure, privacy-preserving, and robust multi-agent decision-making in uncertain and dynamic networked environments. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.