REVS: Unlearning Sensitive Information in Language Models via Rank Editing in the Vocabulary Space

TL;DR

REVS is a new method that unlearns sensitive information from language models by editing neuron ranks in the vocabulary space, improving privacy and robustness without costly data scrubbing.

Contribution

It introduces a non-gradient-based approach for unlearning sensitive data by targeting specific neurons, enhancing privacy preservation in language models.

Findings

REVS outperforms existing methods in unlearning effectiveness.

REVS is more robust against extraction attacks.

REVS maintains model performance after unlearning.

Abstract

Language models (LMs) risk inadvertently memorizing and divulging sensitive or personally identifiable information (PII) seen in training data, causing privacy concerns. Current approaches to address this issue involve costly dataset scrubbing, or model filtering through unlearning and model editing, which can be bypassed through extraction attacks. We propose REVS, a novel non-gradient-based method for unlearning sensitive information from LMs. REVS identifies and modifies a small subset of neurons relevant for constituent tokens that form sensitive information. To adequately evaluate our method on truly sensitive information, we curate three datasets: email and URL datasets naturally memorized by the models, and a synthetic social security number dataset that we tune the models to memorize. Compared to other methods, REVS demonstrates superior performance in unlearning sensitive information and robustness to extraction attacks, while retaining underlying model integrity.