Skip to content
InfoResearchPreprintLLM-specific

Jailbreak Detection on Large Language Models via Contrastive Prototype Structures

Published
Record updated
View JSON

Summary

Researchers propose JDPI, a jailbreak detection method that builds harmful and benign prototypes from binary safety labels and classifies inputs by relative prototype distance, avoiding a separate downstream classifier. Across four benchmarks, contrastive fine-tuning raises detection accuracy by 9.1-18.0 percentage points, and JDPI exceeds the released Llama Guard 3-8B baseline by 0.173-0.310 in ROC-AUC with dataset-specific adaptation. In multi-turn replay, using accumulated user history yields higher conversation-level accuracy than using the current turn alone.