Adversarial Purification with One-Step Guided Diffusion Model

This repository is the official implementation of Adversarial Purification with One-Step Guided Diffusion Model

Requirements

To install requirements:

pip install -r requirements.txt

If you want to enable mixed-precision computation using the apex library, please install the apex library.

git clone https://github.com/NVIDIA/apex
cd apex
pip3 install -v --no-cache-dir ./

(Maybe) Solution for 'IndexError: tuple index out of range': NVIDIA/apex#694 (comment)

Pre-trained Models

You can download pretrained models here:

• DDPM on ImageNet https://github.com/openai/guided-diffusion
  • checkpoint 256x256_diffusion_uncond.pt
• DDPM on Cifar10 https://github.com/openai/improved-diffusion
  • checkpoint cifar10_uncond_50M_500K.pt

Evaluation

To evaluate my model on Cifar10, run:

python3 -m torch.distributed.launch --nproc_per_node=4 run.py --mode pgd --T 400 --scale 50000

To evaluate my model on ImageNet, run:

python3 -m torch.distributed.launch --nproc_per_node=4 run.py --mode pgd --T 100 --scale 2000 --model resnet152

Contributing

This implementation is based on / inspired by:

• https://github.com/ethz-privsec/diffusion_denoised_smoothing
• https://github.com/NVlabs/DiffPure
• https://github.com/openai/guided-diffusion
• https://github.com/openai/improved-diffusion
• https://github.com/point0bar1/ebm-defense