Saved in:
Bibliographic Details
Main Authors: Wei, Boyi, Huang, Kaixuan, Huang, Yangsibo, Xie, Tinghao, Qi, Xiangyu, Xia, Mengzhou, Mittal, Prateek, Wang, Mengdi, Henderson, Peter
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2402.05162
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914987531829248
author Wei, Boyi
Huang, Kaixuan
Huang, Yangsibo
Xie, Tinghao
Qi, Xiangyu
Xia, Mengzhou
Mittal, Prateek
Wang, Mengdi
Henderson, Peter
author_facet Wei, Boyi
Huang, Kaixuan
Huang, Yangsibo
Xie, Tinghao
Qi, Xiangyu
Xia, Mengzhou
Mittal, Prateek
Wang, Mengdi
Henderson, Peter
contents Large language models (LLMs) show inherent brittleness in their safety mechanisms, as evidenced by their susceptibility to jailbreaking and even non-malicious fine-tuning. This study explores this brittleness of safety alignment by leveraging pruning and low-rank modifications. We develop methods to identify critical regions that are vital for safety guardrails, and that are disentangled from utility-relevant regions at both the neuron and rank levels. Surprisingly, the isolated regions we find are sparse, comprising about $3\%$ at the parameter level and $2.5\%$ at the rank level. Removing these regions compromises safety without significantly impacting utility, corroborating the inherent brittleness of the model's safety mechanisms. Moreover, we show that LLMs remain vulnerable to low-cost fine-tuning attacks even when modifications to the safety-critical regions are restricted. These findings underscore the urgent need for more robust safety strategies in LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2402_05162
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
Wei, Boyi
Huang, Kaixuan
Huang, Yangsibo
Xie, Tinghao
Qi, Xiangyu
Xia, Mengzhou
Mittal, Prateek
Wang, Mengdi
Henderson, Peter
Machine Learning
Artificial Intelligence
Computation and Language
Large language models (LLMs) show inherent brittleness in their safety mechanisms, as evidenced by their susceptibility to jailbreaking and even non-malicious fine-tuning. This study explores this brittleness of safety alignment by leveraging pruning and low-rank modifications. We develop methods to identify critical regions that are vital for safety guardrails, and that are disentangled from utility-relevant regions at both the neuron and rank levels. Surprisingly, the isolated regions we find are sparse, comprising about $3\%$ at the parameter level and $2.5\%$ at the rank level. Removing these regions compromises safety without significantly impacting utility, corroborating the inherent brittleness of the model's safety mechanisms. Moreover, we show that LLMs remain vulnerable to low-cost fine-tuning attacks even when modifications to the safety-critical regions are restricted. These findings underscore the urgent need for more robust safety strategies in LLMs.
title Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2402.05162