Machine Unlearning of Pre-trained Large Language Models

TL;DR

This work tackles the ethical and practical challenge of the right to be forgotten for pre-trained LLMs by proposing a unified machine unlearning framework and evaluating seven approximate methods. It introduces a principled objective, derives gradient-based unlearning strategies, and benchmarks on three domains (arXiv, GitHub, books) using an approximate retraining baseline and Membership Inference Attacks. Key findings show that unlearning is orders of magnitude more efficient than retraining, with gradient ascent plus in-distribution descent offering robust hyperparameter behavior and better downstream utility. The study provides detailed guidelines for hyperparameter tuning and contributes open datasets to spur further research in ethical AI and responsible deployment of LLMs. Overall, the work advances practical unlearning techniques for pre-trained LLMs and highlights important considerations for privacy, copyright, and model reliability.

Abstract

This study investigates the concept of the `right to be forgotten' within the context of large language models (LLMs). We explore machine unlearning as a pivotal solution, with a focus on pre-trained models--a notably under-researched area. Our research delineates a comprehensive framework for machine unlearning in pre-trained LLMs, encompassing a critical analysis of seven diverse unlearning methods. Through rigorous evaluation using curated datasets from arXiv, books, and GitHub, we establish a robust benchmark for unlearning performance, demonstrating that these methods are over $10^5$ times more computationally efficient than retraining. Our results show that integrating gradient ascent with gradient descent on in-distribution data improves hyperparameter robustness. We also provide detailed guidelines for efficient hyperparameter tuning in the unlearning process. Our findings advance the discourse on ethical AI practices, offering substantive insights into the mechanics of machine unlearning for pre-trained LLMs and underscoring the potential for responsible AI development.

Figures (3)

• Figure 1: Overview of unlearning pre-trained LLMs to address user removal requests.
• Figure 2: \ref{['fig:1a', 'fig:1b', 'fig:1c']} are visualization of unlearning results on GitHub code across varying learning rates while optimization steps are fixed at 4. \ref{['fig:1d', 'fig:1e', 'fig:1f']} are visualization of unlearning results on GitHub code across varying optimization steps while learning rate is fixed at $2\times 10^{-5}$. In \ref{['fig:1a', 'fig:1d']}, values below 40 are presented on a $\log_{10}$ scale, whereas values above 40 adopt a $\log_{10^{100}}$ scale. The horizontal spring-green line in \ref{['fig:1a', 'fig:1d']} delineates the approximate retraining baseline as the unlearning target. For \ref{['fig:1b', 'fig:1e']}, the scale transitions from $\log_{10}$ for values under 20 to $\log_{10^{10}}$ for values exceeding 20.
• Figure 3: \ref{['fig:2a', 'fig:2b', 'fig:2c']} are visualization of unlearning results on GitHub code across varying batch sizes while other hyperparameters are fixed. \ref{['fig:2d', 'fig:2e', 'fig:2f']} are visualization of unlearning results on GitHub code across varying forget set size while other hyperparameters are fixed. In \ref{['fig:2a', 'fig:2d']}, values below 40 are presented on a $\log_{10}$ scale. Values in \ref{['fig:2a']} above 40 adopt a $\log_{10^{100}}$ scale, whereas values in \ref{['fig:2d']} above 40 adopt a $\log_{10^{10}}$ scale. The horizontal spring-green line in \ref{['fig:2a', 'fig:2d']} delineates the approximate retraining baseline as the unlearning target. For \ref{['fig:2b', 'fig:2e']}, the scale transitions from $\log_{10}$ for values under 20 to $\log_{10^{10}}$ for values exceeding 20.