Advances in Diffusion Models and Unlearning

The field of diffusion models is moving towards increased focus on unlearning and evaluation of these models. Researchers are developing new methods to remove unwanted information from diffusion models, such as Concept Unlearning by Modeling Key Steps of Diffusion Process, which strategically focuses on pivotal steps in the diffusion process to reduce the number of parameter updates needed for effective unlearning. Another area of focus is on understanding how diffusion models learn concepts, with methods like Concept-TRAK providing concept-level attribution to identify contributions to specific elements in generated images. Notable papers in this area include Towards a Principled Evaluation of Knowledge Editors, which highlights the need for more robust evaluation methodologies for knowledge editors, and Image Can Bring Your Memory Back, which proposes a novel adversarial framework to compromise the robustness of unlearned image generation models.

Sources

Towards a Principled Evaluation of Knowledge Editors

Prompt-Free Conditional Diffusion for Multi-object Image Augmentation

Concept Unlearning by Modeling Key Steps of Diffusion Process

Concept-TRAK: Understanding how diffusion models learn concepts through concept-level attribution

Automating Evaluation of Diffusion Model Unlearning with (Vision-) Language Model World Knowledge

Image Can Bring Your Memory Back: A Novel Multi-Modal Guided Attack against Image Generation Model Unlearning

OPC: One-Point-Contraction Unlearning Toward Deep Feature Forgetting

Low Resource Reconstruction Attacks Through Benign Prompts

Built with on top of