Erasing Concepts from Diffusion Models

By Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, David Bau. IEEE International Conference on Computer Vision, 2023

604 citations recorded by the source corpus.

Motivated by concerns that large-scale diffusion models can produce undesirable output such as sexually explicit content or copyrighted artistic styles, we study erasure of specific concepts from diffusion model weights. We propose a fine-tuning method that can erase a visual concept from a pre-trained diffusion model, given only the name of the style and using negative guidance as a teacher. We benchmark our method against previous approaches that remove sexually explicit content and demonstrate its effectiveness, performing on par with Safe Latent Diffusion and censored training. To evaluate artistic style removal, we conduct experiments erasing five modern artists from the network and conduct a user study to assess the human perception of the removed styles. Unlike previous methods, our approach can remove concepts from a diffusion model permanently rather than modifying the output at the inference time, so it cannot be circumvented even if a user has access to model weights. Our code, data, and results are available at erasing.baulab.info.

Subfields: Computer Vision and Pattern Recognition, Signal Processing

View Paper: https://www.semanticscholar.org/paper/7a106b9e32a40b523e80ef1ef262f39213aeed81

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Signal Processing

Citations in the corpus

SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation (2023, background): “While current studies (Gandikota et al., 2023; Zhang et al., 2023a; Heng & Soh, 2023) provide strategies for concept erasure in diffusion models, achieving precision comparable to exact unlearning remains challenging.”
Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models (2023, neutral): “Baselines Our approach is compared against other concurrent works, ESD [17] and ACTD [26].”
Editing Large Language Models: Problems, Methods, and Opportunities (2023, positive): “Moreover, it’s vital for models to forget sensitive knowledge, aligning with machine unlearning concepts (Hase et al., 2023; Wu et al., 2022; Tarun et al., 2021; Gandikota et al., 2023).”
Ablating Concepts in Text-to-Image Diffusion Models (2023, methodology): “[20] aims to remove concepts using a score-based formulation.”
Unified Concept Editing in Diffusion Models (2023, methodology): “Solving for the weight update ∆ W instead of directly solving for W , our method reduces to MEMIT closed-form update.”
Effective Data Augmentation With Diffusion Models (2023, positive): “Given a list of class names in these datasets, we utilize a recent method developed by Gandikota et al. (2023) that fine-tunes the UNet backbone of Stable Diffusion so that concepts specified by a given prompt can no longer be generated (we use class names as such prompts).”
MACE: Mass Concept Erasure in Diffusion Models (2024, methodology): “However, most meth-ods either overlook the residual information of the target phrase embedded within co-existing words, focusing solely on the target phrase [16, 17, 71], or they ﬁnetune uniformly across timesteps [16, 25, 30, 71].”
Rethinking machine unlearning for large language models (2024, neutral): “…of MU has expanded to encompass diverse domains, such as image classification Ginart et al. (2019); Golatkar et al. (2020) In the literature, ‘exact’ unlearning, which involves retraining the model from scratch after removing specific training data points, is often considered the gold standard.”