Constrained Optimization Framework for Unlearning in Diffusion Models
A new framework for unlearning in diffusion models uses constrained optimization to remove undesirable data while preserving model utility. The approach formulates unlearning as minimizing deviation from a pretrained model under separation constraints from unlearning distributions. Three problems are defined using reverse and forward KL divergences and likelihood constraints, with the third offering a novel formulation. Strong duality is established for all problems, enabling explicit characterization of optimal solutions and development of primal-dual algorithms. Experimental results demonstrate effectiveness.
Key facts
- Unlearning in diffusion models aims to remove undesirable data or concepts while preserving pretrained model utility.
- The framework formulates unlearning as minimizing deviation from a pretrained model subject to separation constraints.
- Three constrained optimization problems are based on reverse and forward KL divergences and likelihood constraints.
- The first two generalize existing approaches for concept and data unlearning.
- The third offers a novel and natural formulation for unlearning.
- Strong duality is established for all three problems despite nonconvexity of KL constraints.
- Optimal solutions are characterized as unlearning targets.
- Primal-dual algorithms are developed for each formulation.
Entities
—