Available at: https://digitalcommons.calpoly.edu/theses/3343
Date of Award
6-2026
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Franz Kurfess
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
Search and rescue (SAR) operations require teams to integrate uncertain information under severe time pressure. Large language model (LLM)-based multi-agent systems (MAS) can support decision-making, but they also risk producing hallucinated outputs and processing unreliable or malicious data. This thesis addresses these risks through three linked studies. First, it proposes a modular eight-agent SAR MAS architecture that reflects the structure of real SAR missions by assigning specialized roles to different agents. Second, it introduces a post-hoc probabilistic verification framework that checks LLM agent outputs against a probabilistic knowledge graph built from historical SAR incidents. Third, the thesis examines indirect adversarial threats, including prompt injection, knowledge-base poisoning, and adversarial examples, and evaluates corresponding lightweight defences. The results show that modular specialization, probabilistic grounding, and adversarial defences can improve the reliability and robustness of SAR decision support. The thesis contributes a prototype, evaluation method, and deployment guidance for safety-critical agentic AI.
Award received:
Culminating Experience Completion Fellowship for Spring 2026