Layer-AdvPatcher introduces an innovative defense methodology that protects Large Language Models from malicious jailbreak attacks while maintaining normal performance.

• Implements an unlearning strategy at specific layers of the neural network
• Utilizes self-exposure techniques to identify vulnerable components
• Applies targeted patches to neutralize attack vectors without comprehensive retraining
• Achieves significant security improvements with minimal impact on general performance

This research addresses critical security concerns as LLMs are increasingly deployed in business applications, providing a practical approach to safeguarding AI systems against exploitation while preserving their utility.

Layer-Level Self-Exposure and Patch: Affirmative Token Mitigation for Jailbreak Attack Defense