EvilEdit

This repository contains the code for the paper EvilEdit: Backdooring Text-to-Image Diffusion Models in One Second (ACM MM 2024).

Environment

Step 1: Pull this repository.

git pull https://github.com/haowang-cqu/EvilEdit
cd EvilEdit

Step 2: Create a Conda environment and install PyTorch.

conda create -n eviledit python=3.10
conda activate eviledit
pip3 install torch torchvision

Step 3: Install other dependencies.

pip3 install -r requirements.txt

Quick Start

You can run edit.py to embed the backdoor into stable-diffusion-v1-5. The default trigger is "beautiful cat", and the default backdoor target is "zebra" (check the code for more details).

conda activate eviledit
CUDA_VISIBLE_DEVICES=0 python edit.py

The backdoored U-Net weights will be stored in the models directory. Run the show.ipynb to see the effect of the backdoor.

Citation

@inproceedings{wang2024eviledit,
  title = {EvilEdit: Backdooring Text-to-Image Diffusion Models in One Second},
  author = {Wang, Hao and Guo, Shangwei and He, Jialing and Chen, Kangjie and Zhang, Shudong and Zhang, Tianwei and Xiang, Tao},
  booktitle = {Proceedings of the 32nd ACM International Conference on Multimedia},
  year = {2024},
  pages = {3657–3665},
  doi = {10.1145/3664647.3680689}
}