README

Overview

This project provides implementations for "Sample-specific Noise Injection For Diffusion-Based Adversarial Purification". The experiments are conducted on the CIFAR-10 and ImageNet datasets. The codebase allows for separate execution of attack generation and evaluation phases.

Datasets

CIFAR-10

• Usage: No need to download manually; the dataset will be automatically downloaded and prepared when running the code.

ImageNet

• Description: ImageNet is a large-scale dataset with millions of images across 1,000 classes.
• Required Version: For this project, you need the ImageNet LSVRC 2012 Training/Validation Set.
• Preparation Steps:
  1. Download the ImageNet 2012 training/validation set from the official ImageNet website.
  2. Convert to LMDB Format:
     • Use provided dataset.py to convert the downloaded dataset into LMDB format for efficient data loading.
     • Ensure that the data is organized correctly to match the expected input format of the code.

Diffusion Model

• Description: The project utilizes a diffusion model for adversarial purification and generation tasks.

• Download Link: https://drive.google.com/file/d/16_-Ahc6ImZV5ClUc0vM5Iivf8OJ1VSif/view?usp=sharing https://openaipublic.blob.core.windows.net/diffusion/jul-2021/256x256_diffusion_uncond.pt

  (Note: Please insert the appropriate download link for the diffusion model.)

Running Experiments

Experiments can be run using the main.py script. The codebase is designed to separate the attack generation phase from the evaluation phase.

Example Script

Below is an example of how to run an experiment for generating adversarial examples using PGD+EOT attack:

for seed in $SEED1; do
    python -u main.py --config cifar10.yml \
        --defense_method diffpure \
        --dataset cifar10 \
        --classifier_name cifar10-wideresnet-28-10 \
        --attack_type pgd \
        --seed $seed \
        --batch_size 16 \
        --n_iter 200 \
        --eot 20 \
        --att_max_timesteps 100 \
        --att_num_denoising_steps 20 \
        --att_sampling_method ddpm \
        --num_process_per_node 2 \
        --use_cuda \
        --port 8888
done

Below is an example of how to run an experiment for evaluation:

for seed in $SEED1; do
    python -u main.py --config cifar10.yml \
        --defense_method diffpure \
        --dataset cifar10 \
        --classifier_name cifar10-wideresnet-28-10 \
        --attack_type pgd \
        --seed $seed \
        --batch_size 32 \
        --n_iter 200 \
        --eot 20 \
        --att_max_timesteps 100 \
        --att_num_denoising_steps 20 \
        --att_sampling_method ddpm \
        --def_max_timesteps 100 \
        --def_num_denoising_steps 100 \
        --def_sampling_method ddpm \
        --num_process_per_node 1 \
        --use_cuda \
        --whitebox_defense_eval \
        --port 8888
done

Below is an example of how to run an experiment for generating adversarial examples using PGD+EOT attack on ImageNet:

for seed in $SEED1; do
    python -u main.py --config imagenet.yml \
        --defense_method diffpure \
        --dataset imagenet \
        --classifier_name imagenet-resnet50 \
        --attack_type pgd \
        --seed $seed \
        --batch_size 1 \
        --n_iter 200 \
        --eot 20 \
        --att_max_timesteps 200 \
        --att_num_denoising_steps 10 \
        --att_sampling_method ddpm \
        --num_process_per_node 4 \
        --use_cuda \
        --port 8888
done

The evaluation is the same as CIFAR-10 above. You can reproduce our results using seed: 121, 122, 123.

Evaluation and Attack Phases

• Attack Generation:

  • Before evaluating the defense mechanisms, adversarial examples need to be generated.
  • Use the appropriate attack parameters to create adversarial samples.

• Evaluation Activation:

  • To activate the evaluation phase, include the following arguments in your command:
    • For standard white-box defense evaluation:

      --whitebox_defense_eval

    • For adaptive defense evaluation:

      --adaptive_defense_eval

  • These flags will trigger the evaluation procedures after adversarial examples have been generated.

Dependencies

• Python 3.7 or higher
• PyTorch 2.0.1
• Other dependencies as listed in requirements.txt

Reference

Our codebase is implemented based on the following code, we would like to thank for their great work. [https://github.com/NVlabs/DiffPure.git] [https://github.com/ml-postech/robust-evaluation-of-diffusion-based-purification.git] [https://github.com/ZSHsh98/EPS-AD.git]

License

This project is licensed under the MIT License. See the LICENSE file for details.