RKLD: Reverse KL-Divergence-based Knowledge Distillation for Unlearning Personal Information in Large Language Models

TL;DR

This paper introduces RKLD, a novel knowledge distillation method based on reverse KL-divergence, to improve unlearning of personal data in large language models while maintaining their utility.

Contribution

The paper proposes RKLD, a new unlearning algorithm specifically designed for large language models to better forget personal information without sacrificing model performance.

Findings

RKLD achieves superior forget quality compared to existing methods.

RKLD maintains model utility effectively after unlearning.

Experimental results validate RKLD's effectiveness in unlearning personal data.

Abstract

With the passage of the Right to Be Forgotten (RTBF) regulations and the scaling up of language model training datasets, research on model unlearning in large language models (LLMs) has become more crucial. Before the era of LLMs, machine unlearning research focused mainly on classification tasks in models with small parameters. In these tasks, the content to be forgotten or retained is clear and straightforward. However, as parameter sizes have grown and tasks have become more complex, balancing forget quality and model utility has become more challenging, especially in scenarios involving personal data instead of classification results. Existing methods based on gradient ascent and its variants often struggle with this balance, leading to unintended information loss or partial forgetting. To address this challenge, we propose RKLD, a novel \textbf{R}everse \textbf{KL}-Divergence-based Knowledge \textbf{D}istillation unlearning algorithm for LLMs targeting the unlearning of personal information. Through RKLD, we achieve significant forget quality and effectively maintain the model utility in our experiments.