Reproduce “Score Forgetting Distillation: A Swift, Data-Free Method for Machine Unlearning in Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianqi Chen, Shujian Zhang, Mingyuan Zhou
Score Forgetting Distillation (SFD) introduces a data-free method for machine unlearning in diffusion models, leveraging cross-class score distillation to selectively remove unwanted concepts. The approach achieves high unlearning accuracy and preserves generation quality, enabling one-step sampling that results in up to a 1,000-fold speedup compared to multi-step DDPM baselines.
Key insights
- Cross-class score distillation can effectively guide a generative model to align the score of a "forgetting class" with an "overriding class," achieving targeted unlearning.
- Data-free machine unlearning can be realized by leveraging a frozen pre-trained score network as a teacher, eliminating the need for original training data.
- Integrating unlearning into a score distillation framework allows for simultaneously achieving effective unlearning and rapid one-step generation from diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.