ATTA (Efficient Adversarial Training with Transferable Adversarial Examples)

Code for CVPR'2020 paper: Efficient Adversarial Training with Transferable Adversarial Examples.

Prerequisites

• Python 3.6.3
• Pytorch 1.2.0, torchvision 0.4.0
• Numpy 1.13.3

Code Overview

The directory models contains model architecture definition files. The directory data-config contains different config files to train the model and the directory data-model is used to contain model checkpoints.

Other seven Python scripts are used to train and evaluate the ATTA model.

• train_atta_mnist.py: trains ATTA models on MNIST dataset.
• train_atta_cifar.py: trains ATTA models on CIFAR10 dataset.
• cifar_dataloader.py: loads the padded data of CIFAR10.
• adv_attack.py: generates accumulative adversarial examples for the training of ATTA.
• adaptive_data_aug.py: performs data augmentation and inverse data augmentation for ATTA.
• pgd_attack_mnist.py: performs PGD-k attack on MNIST models.
• pgd_attack_cifar10.py: performs PGD-k attack on CIFAR10 models.

Simple instructions to train and evaluate models

Train a model:

#MNIST

python train_atta_mnist.py --config-file [config_file_name] --gpuid [GPU_ID]


#CIFAR10

python train_atta_cifar.py --config-file [config_file_name] --gpuid [GPU_ID]

Attack a model:

#MNIST

python pgd_attack_mnist.py --model-dir [path of model] --gpuid [GPU_ID]


#CIFAR10

python pgd_attack_cifar10.py --model-dir [path of model] --gpuid [GPU_ID]

Naming rule for configuration files in data-config:

[dataset]-atta-[the number of attack iterations]-[training method].json

data-config/mnist-atta-1-mat.json means that model will be trained with MAT(ATTA-1) on MNIST.

Examples for training and evaluate:

• TRADES(ATTA-1) on MNIST:

python train_atta_mnist.py --config-file data-config/mnist-atta-1-trades.json --gpuid 0

python pgd_attack_mnist.py --model-dir data-model/mnist-trades-atta-1/model-mnist-epoch60.pt --gpuid 0

• MAT(ATTA-1) on CIFAR10:

python train_atta_cifar.py --config-file data-config/cifar-atta-1-mat.json --gpuid 0

python pgd_attack_cifar10.py --model-dir data-model/cifar-mat-atta-1/model-cifar-epoch38.pt --gpuid 0

Reference

@inproceedings{zheng2020efficient,
 author={Zheng, Haizhong and Zhang, Ziqi and Gu, Juncheng and Lee, Honglak and Prakash, Atul},
 title={Efficient Adversarial Training with Transferable Adversarial Examples},
 BOOKTITLE = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
 YEAR = {2020}
}

Credit: The implementation of ATTA is based on TRADES code.