[ECCV 2024 - Oral] Adversarial Robustification via Text-to-Image Diffusion Models

This repository contains code for the paper Adversarial Robustification via Text-to-Image Diffusion Models by Daewon Choi*, Jongheon Jeong*, Huiwon Jang and Jinwoo Shin. (*Equally Contributed)

In this paper, by leveraging text-to-image diffusion models, we provide a scalable framework that does not require any external datasets in robustifiying image classifiers. We successfully incoporate text-to-image diffusion models into denoisy-classify pipeline as adaptable framework that can be self-optimized to specific tasks.

📣 News

• [2024.08.17] Code is uploaded.
• [2024.08.12] Our paper is accepted for Oral presentation in ECCV 2024 🎉.
• [2024.07.01] Our paper is accepted by ECCV 2024 🎉.

💻 Get Started

Set Environments

This repository already contain diffusers 0.20.0 files. So, you don't need to install diffusers. Just track below environment setting.

# git clone this repository
git clone https://github.com/ChoiDae1/robustify-T2I.git
cd robustifiy-T2I

# make conda virtual environment for python version
conda create -n certified python=3.10.12 
conda activate certified

# make python virtual environment 
python -m venv .venv 
source .venv/bin/activate
conda deactivate

# download necessary libraries
pip install -r requirements.txt 

You should replace files in .venv/lib/python3.10.12/site-packages/torchvision/datasets with the files in replace folder in this repo.

Self-Adaptation

In this phase, we first synthesize a few reference images via text-to-image diffusion models and then, fine-tune text-to-image diffusion models and classifiers with them. Other adaptation details (ex. hyper-parameters) is reported in Appendix in our paper. By executing below code sequentially, you finally obtain checkpoints.

# synthesize a few reference images
sh scripts/synthesize_set.sh

# self-personalization 
sh scripts/self_personalization.sh 

# classifier fine-tuning 
sh scripts/classifier_finetuning.sh

Evaluation

For evaluation, we provide two type of codes, one is for testing empirical accuracy and other is for certified accuracy. After code execution, output file is saved at path --output_dir. Other evaluation details (ex. attack/smoothing parameters) is reported in Appendix in our paper.

# for empirical accuracy
sh scripts/eval_empirical.sh

# for certified accuracy 
sh scripts/eval_certified.sh

After obtaining the output file, you can visualize/calculate accuracy by using jupyter file analysis/analyze.py. We provide some examples in analyze_example.ipynb. When computing empirical clean accuracy, you directly execute measure_empirical_clean.py by using output file and classifier checkpoint.

📝 Citation

@inproceedings{RobustifyT2I,
  title={Adversarial Robustification via Text-to-Image Diffusion Model},
  author={Daewon Choi and Jongheon Jeong and Huiwon Jang and Jinwoo Shin},
  booktitle={European Conference on Computer Vision},
  year={2024}
}

Acknowledgment

This project is based on randomzied smoothing. We also thank diffusers, clip for the pretrained models, TeCoA for dataset settings.