Towards Assurance of LLM Adversarial Robustness using Ontology-Driven Argumentation

A novel approach aims to ensure large language models' robustness against adversarial attacks by using formal argumentation with ontologies. This method structures state-of-the-art attacks and defenses, creating human-readable assurance cases and machine-readable representations. Targeting engineers, data scientists, users, and auditors, it addresses challenges in managing the implicit and heterogeneous knowledge required for continuous robustness assurance in LLMs. Demonstrated in English language and code translation tasks, the approach emphasizes the importance of formalizing heterogeneous knowledge for complex AI technologies, particularly in security, to improve confidence in system quality.