Reliable Unlearning Harmful Information in LLMs with Metamorphosis Representation Projection

TL;DR

This paper introduces a novel method called Metamorphosis Representation Projection (MRP) that enhances the unlearning of harmful information in Large Language Models by applying irreversible projections, improving safety and robustness.

Contribution

The paper proposes MRP, a new approach using irreversible projection in hidden states to effectively eliminate harmful data and defend against relearning attacks in LLMs.

Findings

MRP achieves state-of-the-art unlearning effectiveness.

MRP effectively defends against relearning attacks.

The method preserves model performance on useful knowledge.

Abstract

While Large Language Models (LLMs) have demonstrated impressive performance in various domains and tasks, concerns about their safety are becoming increasingly severe. In particular, since models may store unsafe knowledge internally, machine unlearning has emerged as a representative paradigm to ensure model safety. Existing approaches employ various training techniques, such as gradient ascent and negative preference optimization, in attempts to eliminate the influence of undesired data on target models. However, these methods merely suppress the activation of undesired data through parametric training without completely eradicating its informational traces within the model. This fundamental limitation makes it difficult to achieve effective continuous unlearning, rendering these methods vulnerable to relearning attacks. To overcome these challenges, we propose a Metamorphosis Representation Projection (MRP) approach that pioneers the application of irreversible projection properties to machine unlearning. By implementing projective transformations in the hidden state space of specific network layers, our method effectively eliminates harmful information while preserving useful knowledge. Experimental results demonstrate that our approach enables effective continuous unlearning and successfully defends against relearning attacks, achieving state-of-the-art performance in unlearning effectiveness while preserving natural performance. Our code is available in https://github.com/ChengcanWu/MRP.