CluCERT: Certifying LLM Robustness via Clustering-Guided Denoising Smoothing

arXiv cs.AI · 6 hours ago

Researchers introduced CluCERT, a framework for certifying Large Language Model robustness against adversarial attacks using clustering-guided denoising smoothing techniques. The method achieved tighter certified robustness bounds compared to existing approaches while reducing computational costs through semantic clustering filters and accelerated synonym substitution strategies. CluCERT enables more efficient evaluation of LLM vulnerability to meaning-preserving input perturbations across various tasks and jailbreak defense scenarios.