Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
Keywords: Alignment, Safety, Large Language Model, Interpretability
TL;DR: We identify safety-critical neurons and ranks in Llama2-chat models, uncovering a sparse 3% structure. Removing these regions effectively compromises safety but retains utility, elucidating the brittleness of safety alignment.
Abstract: 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.
Submission Number: 51
Loading