Q-resafe: Assessing Safety Risks and Quantization-aware Safety Patching for Quantized Large Language Models

TL;DR

This paper evaluates the safety risks introduced by quantization in large language models and proposes Q-resafe, a framework to restore safety capabilities with minimal utility loss, validated through extensive experiments.

Contribution

It provides the first comprehensive safety assessment of quantized LLMs and introduces a novel safety patching framework tailored for quantization-induced vulnerabilities.

Findings

Q-resafe effectively restores safety to quantized LLMs.

Quantization can significantly impair LLM safety capabilities.

The framework maintains model utility while improving safety.

Abstract

Quantized large language models (LLMs) have gained increasing attention and significance for enabling deployment in resource-constrained environments. However, emerging studies on a few calibration dataset-free quantization methods suggest that quantization may compromise the safety capabilities of LLMs, underscoring the urgent need for systematic safety evaluations and effective mitigation strategies. In this paper, we present comprehensive safety evaluations across various mainstream quantization techniques and diverse calibration datasets, utilizing widely accepted safety benchmarks. To address the identified safety vulnerabilities, we propose a quantization-aware safety patching framework, Q-resafe, to efficiently restore the safety capabilities of quantized LLMs while minimizing any adverse impact on utility. Extensive experimental results demonstrate that Q-resafe successfully re-aligns the safety of quantized LLMs with their pre-quantization counterparts, even under challenging evaluation scenarios. Project page is available at: https://github.com/Thecommonirin/Qresafe.