EVA: Editing for Versatile Alignment against Jailbreaks.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Natural Language Processing

Background:

• Large Language Models (LLMs) and Vision Language Models (VLMs) exhibit advanced capabilities but are susceptible to jailbreaking attacks.
• Existing defenses like safety fine-tuning or external filters can be computationally expensive and compromise model utility.
• These methods often lead to a safety-utility trade-off, negatively impacting performance on benign tasks.

Purpose of the Study:

• To introduce EVA (Editing for Versatile Alignment against Jailbreaks), a novel framework for safety alignment using direct model editing.
• To address the limitations of current safety measures in LLMs and VLMs, specifically computational overhead and the safety-utility trade-off.
• To provide a precise and efficient solution for post-deployment safety alignment against jailbreaking attacks.

Main Methods:

• EVA reframes safety alignment as a knowledge correction task, focusing on precise edits rather than extensive retraining.
• The framework identifies and surgically edits specific neurons responsible for susceptibility to harmful instructions.
• It ensures that the vast majority of the model's parameters remain unchanged, preserving general reasoning capabilities.

Main Results:

• EVA effectively neutralizes harmful behaviors by localizing updates to specific neurons.
• Experiments show EVA outperforms existing methods in mitigating jailbreaks across both LLMs and VLMs.
• The proposed method demonstrates a precise and efficient approach to safety alignment without compromising model performance.

Conclusions:

• EVA offers a groundbreaking approach to safety alignment in LLMs and VLMs through direct model editing.
• The framework successfully mitigates jailbreaking vulnerabilities while preserving the model's utility and general capabilities.
• EVA presents a promising, efficient, and precise solution for enhancing AI safety in deployed models.