Pytorch Implementation of Adversarial Training with Projected Gradient Descent and Fast Gradient Sign Methods

This reopository is made by Byungjoo Kim, Korea University.

Installation

1. Create your virtual environments ('conda' recommended).

conda create -n ${envname} python=3.6

Activate the virtual environment. source activate ssp

2. Install require packages, this codes require PyTorch 1.1, overrides, tqdm. For PyTorch, follow this website(note that you must install PyTorch>=1.1).

Make sure that this repository uses Tensorboard, you should install tensorflow. In my machine, I used CUDA 9.0, so I installed tensorflow 1.9.0.

pip install tensorflow-gpu==1.9.0
pip install overrides
pip install https://github.com/bethgelab/foolbox/archive/master.zip
pip install tqdm
pip install tb-nightly
pip install future

3. Now, clone this repository.

git clone https://github.com/matbambbang/pgd_adversarial_training.git

This would make the repository named pgd_adversarial_training. Move to this repository, and follow below descriptions.

Description

This module is build for MNIST and CIFAR10. For each task, I implemented different networks. Please see model/mnist.py and model/cifar10.py.

Training

I would upload the sample trained model soon.

For train the model, use cifar10_train.py.

python mnist_train.py --model ${model} --block ${block} --save ${save_path} --norm ${norm} --tbsize ${batch_size} --adv ${adv}
python cifar10_train.py --model ${model} --block ${block} --save ${save_path} --norm ${norm} --tbsize ${batch_size} --adv ${adv}

When train network while each groups are composed of 6 residual blocks,

python mnist_train.py --model res --block 6 --save res_mnist_6
python cifar10_train.py --model res --block 6 --save res_cifar_6

In the table below, we describe the arguments that you can control:

args	Valid arguments
model	res(default), wres, conv, in MNIST, wres is not available
block	6(default), you can use any integer values. In CIFAR10, this argument only effects on res or wres
epochs	200(default) , in MNIST, 100 is default
lr	0.1(default), 0.001 or 0.0001 recommended when you use adam optimizer
decay	0.0005(default), not available when using adam optimizer
opt	sgd(default) or adam, when using GroupNorm, adam is recommended
norm	b(BatchNorm, default), g
tbsize	128(default), you can use any integer values
adv	none(default), for adversarial training, use fgsm, pgd, or ball
save	identify the folder name in this arguments, I recommend you the name should be the combination of model, block and norm, i.e., res_6_b

After training your model, the models and logs are saved in experiments/${save_path}.

Attack the trained model

After the model is trained, you can attack your model with attack algorithms. This repository supports l_inf bounded attack, with using sign of gradients.

Valid attacks : fgsm, pgd.

python attack_test.py --model ${model} --eval ${eval} --attack ${attack} --eps ${eps} --load ${load} --norm ${norm}

args	Valid arguments
attack	fgsm(default), bim, mim, pgd, ball
eval	cifar10(default), mnist
eps	8.0(default), any integer pixel intensity values. When the model is for MNIST classification, eps should
load	Identify the folder name which includes target model