Safety Alignment via Constrained Knowledge Unlearning

TL;DR

This paper introduces Constrained Knowledge Unlearning (CKU), a novel method for improving safety in large language models by selectively unlearning harmful knowledge while preserving useful information, thereby enhancing safety without sacrificing performance.

Contribution

The paper proposes CKU, a new safety alignment technique that localizes and unlearns harmful knowledge in LLMs through neuron scoring and gradient pruning, improving safety and interpretability.

Findings

CKU significantly reduces harmful outputs in LLMs.

CKU maintains overall model performance while enhancing safety.

Neuron analysis reveals insights into safety and knowledge retention.

Abstract

Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks. Existing defense mechanisms have not fully deleted harmful knowledge in LLMs, which allows such attacks to bypass safeguards and produce harmful outputs. To address this challenge, we propose a novel safety alignment strategy, Constrained Knowledge Unlearning (CKU), which focuses on two primary objectives: knowledge localization and retention, and unlearning harmful knowledge. CKU works by scoring neurons in specific multilayer perceptron (MLP) layers to identify a subset U of neurons associated with useful knowledge. During the unlearning process, CKU prunes the gradients of neurons in U to preserve valuable knowledge while effectively mitigating harmful content. Experimental results demonstrate that CKU significantly enhances model safety without compromising overall performance, offering a superior balance between safety and utility compared to existing methods. Additionally, our analysis of neuron knowledge sensitivity across various MLP layers provides valuable insights into the mechanics of safety alignment and model knowledge editing.