Adversarial Diffusion Bridge Model

This is the code for ADBM: Adversarial Diffusion Bridge Model for Reliable Adversarial Purification.

Taking CIFAR-10 as an example, the following commands can be used to train and test the ADBM model.

Training ADBM on CIFAR-10

The required model ``32x32_ckpt_8_diffusion.pth'' can be downloaded from this link provided by the DiffPure repo.

accelerate launch --multi_gpu --num_processes 4 main.py --configs ft_fz_wcls_cifar10_sde.yaml

Test ADBM on CIFAR-10 with PGD-200 Linf attack (EOT and full gradient)

python evaluate.py --gpu 4 \
                --scheduler ddim \
                --steps 5 \
                --noise_level 100 \
                --stack 1 \
                --configs configs/ft_fz_wcls_cifar10_sde.yaml \
                --attack_method PGD \
                --attack_type Linf \
                --unet_ckpt /dir/to/trained/ADBM/UNet/ckpt

Requirements

• Python 3.8.16
• PyTorch 1.12.1
• CUDA 11.3
• numpy
• accelerate

Acknowledgement

If you find that our work is helpful to you, please star this project and consider cite:

@inproceedings{li2024adbm,
  author    = {Li, Xiao and Sun, Wenxuan and Chen, Huanran and Li, Qiongxiu and Liu, Yining and He, Yingzhe and Shi, Jie and Hu, Xiaolin},
  title     = {ADBM: Adversarial diffusion bridge model for reliable adversarial purification},
  booktitle = International Conference on Learning Representations (ICLR),
  year      = {2025}
}