Date of Submission

2-26-2026

Document Type

Dissertation

Department

Engineering and Applied Science Education

Advisor

Vahid Behzadan, Ph.D

Committee Member

Shayok Mukhopadhyay, Ph.D.

Committee Member

Khaled Sayed, Ph.D.

Committee Member

Stefan Sarkadi, Ph.D.

LC Subject Headings

Multiagent systems, Deception, Reinforcement learning, Game theory, Computer Security, Intelligent agents (Computer software)

Abstract

In multi-agent learning systems where agents interact in dynamic and complex environments, cooperation, coordination and communication are fundamental to achieving collective or individual goals. However, the rise of both emergent and induced deceptive behaviors pose significant challenges to system integrity and performance. Emergent deception and deceptive behaviors may manifest as a result of objective misalignment or incomplete situational awareness, leading to systematic failure in coordination. Agents that engage in intentional deceptive behaviors manipulate outcomes for their self-interest, undermining the fairness, security and reliability of these systems. Modeling and mitigating these forms of deception is critical to not only ensure effective cooperation, but also ensure the trustworthiness of multi-agent systems in real-world applications like autonomous vehicle coordination, algorithmic trading, and distributed control of critical infrastructure. Characterizing and mitigating deception are essential for the deployment of multi-agent systems that are aligned with human ethical standards, intentions, and system-level goals. Multi-Agent Reinforcement Learning (MARL) frameworks enable agents to learn complex behaviors through interaction, but this learning process can produce unintended consequences. In resource-constrained or competitive environments, agents may develop deceptive strategies that appear cooperative while actually serving misaligned or malicious goals. Although prior research has shown that agents can learn deceptive strategies, a comprehensive framework for analyzing and characterizing deceptive agents in Multi-Agent Systems (MAS) is still lacking. This research addresses this gap by creating a formal framework to describe, detect, and mitigate deception in both learning and rule-based MAS. Leveraging principles from game theory, information theory, and machine learning security, the framework aims to ensure the secure and practical deployment of MAS, particularly in critical settings and infrastructure.

Download

Share

COinS