Safety Layers in Aligned Large Language Models: The Key to LLM Security

TL;DR

This paper identifies specific internal layers in aligned large language models that are crucial for security, and proposes a fine-tuning method that preserves security without sacrificing performance.

Contribution

It uncovers the existence and location of safety layers within LLMs and introduces SPPFT, a fine-tuning technique that maintains security during model updates.

Findings

Safety layers are located in the middle of the model.

SPPFT preserves security and performance during fine-tuning.

The method reduces computational resources needed for fine-tuning.

Abstract

Aligned LLMs are secure, capable of recognizing and refusing to answer malicious questions. However, the role of internal parameters in maintaining such security is not well understood yet, further these models can be vulnerable to security degradation when subjected to fine-tuning attacks. To address these challenges, our work uncovers the mechanism behind security in aligned LLMs at the parameter level, identifying a small set of contiguous layers in the middle of the model that are crucial for distinguishing malicious queries from normal ones, referred to as ``safety layers". We first confirm the existence of these safety layers by analyzing variations in input vectors within the model's internal layers. Additionally, we leverage the over-rejection phenomenon and parameters scaling analysis to precisely locate the safety layers. Building on these findings, we propose a novel fine-tuning approach, Safely Partial-Parameter Fine-Tuning (SPPFT), that fixes the gradient of the safety layers during fine-tuning to address the security degradation. Our experiments demonstrate that the proposed approach can significantly preserve LLM security while maintaining performance and reducing computational resources compared to full fine-tuning.