UNLEARNING IS BETTER THAN UNSEEN: UNLEARNING SCORE-BASED GENERATIVE MODEL
Keywords: Machine Unlearning，Score-based generative model
TL;DR: We investigate the generative unlearning and propose the first unlearning score-based generative model.
Abstract: Diffusion generative models, including Score-Based Generative Models (SGM) and Denoising Diffusion Probabilistic Models (DDPM), have demonstrated remarkable performance across various domains in recent years. However, concerns regarding privacy and potential misuse of AI-generated content have become increasingly prominent. While generative unlearning methods have been investigated on DDPM models, research on unlearning SGM is still largely missing. Furthermore, the current 'gold standard' of machine unlearning---retraining a model from scratch after removing the undesirable data, does not perform well in SGM and its downstream tasks, such as image inpainting and reconstruction. To fill this gap, we propose the first Score-based Generative Unlearning (SGU) for SGM, which surpasses the previous 'gold standard' of unlearning.SGU introduces a new score adjustment strategy that deviates the learned score from the original undesirable data score during the continuous-time stochastic differential equation process. Extensive experimental results demonstrate that SGU significantly reduces the likelihood of generating undesirable content while preserving high quality for normal image generation. Albeit designed for SGM, SGU is a general and flexible unlearning framework that is compatible with diverse diffusion architectures (SGM and DDPM) and training strategies (re-training and fine-tuning), and enables zero-shot transfer of the unlearning generative model to downstream tasks, including image inpainting and reconstruction. The code will be shared upon acceptance.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1424
Loading