/TPAP

Primary LanguagePython

Robust Overfitting Does Matter: Test-Time Adversarial Purification With FGSM (CVPR 2024)

Introduction

  • 下面是这篇文章的代码，主要是用 PyTorch 编写的。
    Here's the code for the article, written primarily in PyTorch.

Usage

Data Preparation


1、加载数据集时，如果数据集不存在，将下述的代码中download=False修改为download=True。
When loading a dataset, if the dataset does not exist, change download=False to download=True in the following code.

CIFAR10(root='./cifar10/', train=True, download=False, transform=transform_train)
CIFAR10(root='./cifar10/', train=False, download=False, transform=transform_test)


2、使用Tiny-ImageNet数据集前参考下面链接，处理val数据。
Refer to the link below before using the Tiny-ImageNet dataset to process the val data.

https://github.com/tly18/TPAP/blob/main/TPAP_tiny_imagenet/val_data.py


在pytorch环境下的linux终端运行文件：
Run the file in a linux terminal in a pytorch environment:

python val_data.py

Environment Configuration

  • Pytorch 环境包括 python 3.8.13、pytorch 1.7.1、CUDA 9.2、torchvision 0.8.2、advertorch 0.2.3、torchattacks 3.3.0 和 matplotlib 3.2.2。
    The pytorch environment includes python 3.8.13, pytorch 1.7.1, CUDA 9.2, torchvision 0.8.2, advertorch 0.2.3, torchattacks 3.3.0, matplotlib 3.2.2.

  • 下面是在线使用 conda 和 pip 配置环境的简单示例。
    The following is a simple example of configuring an environment online using conda and pip.

  • linux环境下安装自己的工作环境 python+pytorch+torchvision。
    Install your own working environment python+pytorch+torchvision in linux environment.


1、创建环境
Create environment

在linux终端输入：
Type in the linux terminal:

conda create -n tly18 python=3.8.13


2、安装部署
Deployment 

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=9.2 -c pytorch


3、安装攻击库
Installation of attack libraries

在linux终端输入：
Type in the linux terminal:

pip install torchattacks
pip install advertorch=0.2.3 -i https://pypi.tuna.tsinghua.edu.cn/simple/

Instructions for the use of FWA


参考FWA方法的使用说明，网址链接如下：
Refer to the FWA method for instructions on how to use the URL link below:

https://github.com/tly18/TPAP/blob/main/FWA/README.md

Train and Test

1、我们提供了CIFAR-10、CIFAR-100、SVHN、TinyImageNet四个数据集的TPAP训练方法。
We provide TPAP training methods for four datasets, CIFAR-10, CIFAR-100, SVHN, and TinyImageNet.

  • 除CIFAR-10在./TPAP目录运行外，CIFAR-100、SVHN、TinyImageNet数据集需要进入各自的目录进行训练。
    With the exception of CIFAR-10 which is run in the . /TPAP directory to run, the CIFAR-100, SVHN, and TinyImageNet datasets require access to their respective directories for training.

  • 例如，训练CIFAR-10：
    For example, training CIFAR-10:

python train_fgsm_at_cifar10_TPAP.py


2、与Trades、MART方法结合时，需要在文件中修改训练的损失函数。
When combined with the Trades, MART method, the loss function for training needs to be modified in the file.

3、我们提供了PGD对抗训练的代码，运行：
We provide code for PGD adversarial training that runs:

python train_pgd_at_cifar10.py


4、我们提供了TPAP在其它攻击方法下的测试代码，运行：
We provide test code for TPAP running under other attack methods:

python TEST_other_attack.py

Results


参考文章中提供的结果：
Refer to the results provided in the article, article link:

https://arxiv.org/abs/2403.11448

Contact


If you have any problem about our code, feel free to contact us.
Lei Zhang (leizhang@cqu.edu.cn)
LinYu Tang (linyutang@cqu.edu.cn or 2367300330@qq.com)
or describe it in Issues.

Citation

@article{tang2024robust,
  title={Robust Overfitting Does Matter: Test-Time Adversarial Purification With FGSM},
  author={Tang, Linyu and Zhang, Lei},
  journal={arXiv preprint arXiv:2403.11448},
  year={2024}
}