Adversarial Training in PyTorch

This is an implementation of adversarial training using the Fast Gradient Sign Method (FGSM) [1], Projected Gradient Descent (PGD) [2], and Momentum Iterative FGSM (MI-FGSM) [3] attacks to generate adversarial examples. The model employed to compute adversarial examples is WideResNet-28-10 [4]. An implementation of this model is retrieved from [5]. The dataset used to conduct the experiment is CIFAR-10.

Usage

Installation

The training environment (PyTorch and dependencies) can be installed as follows:

git clone https://github.com/AlbertMillan/adversarial-training-pytorch.git
python setup.py install

Tested under Python 3.8.0 and PyTorch 1.4.0

Arguments

This model offers a significant degree of customization. The following are the list of arguments:

Storage Variables

Command	Default Value	Description
--ds_path	'datasets/'	Path to dataset.
--load_dir	'chkpt/chkpt_plain/'	Path to pre-trained model. Used to generate adversarial examples from the test set.
--load_name	'chkpt__model_best.pth.tar'
--load_adv_dir	'chkpt/chkpt_plain/'
--load_adv_name	'chkpt__model_best.pth.tar'	File name
--save_dir	'chkpt/new/'	Path to store model checkpoints on each iteration.

Model Hyper-parameters

Command	Default Value	Description
--lr	0.1	Learning rate.
--itr	76	Number of training iterations.
--batch_size	64	Batch size.
--momentum	0.9	Momentum constant.
--nesterov	True	Whether to apply Nesterov momentum.
--weight_decay	2e-4	Weight decay.
--topk	1	Compute accuracy over top k-predictions

Adversarial Generator Properties

Command	Default Value	Description
--eps	(8./255.)	Epsilon (float)
--attack	0	Attack type (0: no-attack; 1: PGD)
--adv_momentum	None	Momentum constant used to generate adversarial examples if given (float).
--train_max_iter	1	Iterations performed to generate adversarial examples from train set.
--test_max_iter	0	Iterations performed to generate adversarial examples from test set.
--train_mode	0	Training on raw images (0), adversarial images (1) or both (2).
--test_mode	0	Testing on raw images (0), adversarial images (1) or both (2).

Other Properties

Command	Default Value	Description
--gpu	"0,1"	Epsilon
--zero_norm	False	Whether to perform zero-mean normalization on the dataset.
--skip_train	False	Wether to perform testing without training, loading pre-trained model.

Setup

Examples of Use

Acknowledgements