OPTML-Group/AdvUnlearn
Official implementation of NeurIPS'24 paper "Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models". This work adversarially unlearns the text encoder to enhance the robustness of unlearned DMs against adversarial prompt attacks and achieves a better balance between unlearning performance and image generation
This project helps AI developers and researchers build safer image generation models by removing unwanted or sensitive content. It takes a pre-trained text-to-image diffusion model and specific concepts (like nudity, certain art styles, or objects) you want to remove. It produces a modified diffusion model that no longer generates those concepts, even when faced with tricky, 'adversarial' prompts. This is for machine learning engineers, AI safety researchers, or MLOps specialists working with generative AI.
No commits in the last 6 months.
Use this if you need to make your text-to-image diffusion models more robust and secure by permanently erasing undesirable concepts, preventing them from being generated even under adversarial text attacks.
Not ideal if you are looking to remove concepts from other types of AI models or if you need a non-developer-friendly interface for concept removal.
Stars
49
Forks
3
Language
Jupyter Notebook
License
CC-BY-4.0
Category
Last pushed
Nov 04, 2024
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/diffusion/OPTML-Group/AdvUnlearn"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
OPTML-Group/Unlearn-Saliency
[ICLR24 (Spotlight)] "SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in...
Shilin-LU/VINE
[ICLR 2025] "Robust Watermarking Using Generative Priors Against Image Editing: From...
WindVChen/DiffAttack
An unrestricted attack based on diffusion models that can achieve both good transferability and...
koninik/DiffusionPen
Official PyTorch Implementation of "DiffusionPen: Towards Controlling the Style of Handwritten...
Wuyxin/DISC
(ICML 2023) Discover and Cure: Concept-aware Mitigation of Spurious Correlation