This research introduces a systematic approach to evaluate fairness in Large Language Models (LLMs) through metamorphic testing, revealing hidden biases especially at intersections of protected attributes.

Key Findings:

• Developed fairness-oriented metamorphic relations to test bias in LLMs
• Identified significant biases in both LLaMA and GPT models
• Discovered intersectional biases that affect multiple demographic groups simultaneously
• Demonstrated heightened risks in sensitive domains like healthcare and law

Why It Matters: As LLMs continue to be deployed in critical security contexts, understanding and mitigating these biases is essential for building trustworthy AI systems that treat all users fairly and prevent potential discrimination.

Metamorphic Testing for Fairness Evaluation in Large Language Models: Identifying Intersectional Bias in LLaMA and GPT