DUET: Distilled LLM Unlearning from an Efficiently Contextualized Teacher

TL;DR

DUET is a distillation-based unlearning method for large language models that effectively removes undesirable knowledge while maintaining utility, offering higher efficiency and robustness against attacks compared to existing approaches.

Contribution

It introduces DUET, a novel unlearning approach combining in-contextualized and distillation techniques to improve efficiency and robustness in LLM unlearning.

Findings

DUET outperforms existing methods in forgetting and utility preservation.

It is significantly more data-efficient than prior unlearning techniques.

Extensive evaluations show DUET's robustness against prompt removal and reverse engineering.

Abstract

LLM unlearning is a technique to remove the impacts of undesirable knowledge from the model without retraining from scratch, which is indispensable towards trustworthy AI. Existing unlearning methods face significant limitations: conventional tuning-based unlearning is computationally heavy and prone to catastrophic forgetting. In contrast, in-contextualized unlearning is lightweight for precise unlearning but vulnerable to prompt removal or reverse engineering attacks. In response, we propose Distilled Unlearning from an Efficient Teacher (DUET), a novel distillation-based unlearning method that combines the merits of these two lines of work. It learns a student model to imitate the behavior of a prompt-steered teacher that effectively refuses undesirable knowledge generation while preserving general domain knowledge. Extensive evaluations on existing benchmarks with our enriched evaluation protocols demonstrate that DUET achieves higher performance in both forgetting and utility preservation, while being orders of magnitude more data-efficient than state-of-the-art unlearning methods.