Robustness, Privacy, and Generalization of Adversarial Training

This repository contains the PyTorch source code for technical report, "Robustness, Privacy, and Generalization of Adversarial Training," by Fengxiang He, Shaopeng Fu, Bohan Wang, and Dacheng Tao.

Code Structures

|----./
    |---- models/
        |---- __init__.py
        |---- resnet.py
    |---- utils/
        |---- datasets/
            |---- __init__.py
            |---- torchvision_datasets.py
        |---- __init__.py
        |---- argument.py
        |---- attack.py
        |---- data.py
        |---- generic.py
    |---- eval.py
    |---- eval_grad.py
    |---- train.py

Requirements

• Python 3.7.2
• pytorch 1.6.0
• torchvision 0.7.0
• numpy 1.19.1

Instructions

The following four scripts are the example scripts for adversarial training, evaluation, and calculating gradient noises:

template_train.sh
template_eval.sh
template_eval_grad_mu.sh
template_eval_grad_noise.sh

We then introduce the usages of these scripts as examples. Noting that you should modify these scripts for your own use.

Training

Run the example script as follow:

bash template_train.sh

After finishing training, the log will be saved in ./temp/log.txt, the data records will be saved in ./temp/ckpts/ckpt-80000-log.pkl, and the trained model will be saved in ./temp/ckpts/ckpt-80000-model.pkl.

Some of the key hyperparameters in template_train.sh is introduced in the following table:

Hyperparameter	Description	Type
--dataset	The choice of dataset	string, should be: cifar10 or cifar100
--calc-mg-freq	The number of interval steps of calculating max gradient norm for robustified intensity	int
--pgd-radius	The radius $\rho$ in PGD	float
--pgd-steps	The number of iterative modification in PGD	int
--pgd-step-size	The step size $\alpha$ in PGD	float
--pgd-norm-type	The norm of metric in PGD	string, should be: l-infty or l2

Evaluation

Run the example script as follow:

bash template_eval.sh

The evaluation results will be saved in ./temp/eval/eval-rad-10-ts-80000.pkl, which includes the accuracies on training / test / adversarial training / adversarial test sets, and the membership inference attack accuracy.

Calculation of Gradient Noises

Step 1: Calculate the gradient on the full training set

Run the example script as follow:

bash template_eval_grad_mu.sh

The gradient on the full training set will be saved in ./temp/eval-grad/ckpt-80000-mu.pkl.

Step 2: Calculate the gradient noises

Run the example script as follow:

bash template_eval_grad_noise.sh

The collected gradient noises will be saved in ./temp/eval-grad/ckpt-80000-noise.pkl.

Citation

@article{he2020robustness,
  title={Robustness, Privacy, and Generalization of Adversarial Training},
  author={He, Fengxiang and Fu, Shaopeng and Wang, Bohan and Tao, Dacheng},
  journal={arXiv preprint arXiv:2012.13573},
  year={2020}
}

Contact

For any issues please kindly contact

Fengxiang He, fengxiang.f.he@gmail.com
Shaopeng Fu, fshp971@gmail.com
Bohan Wang, bhwangfy@gmail.com
Dacheng Tao, dacheng.tao@gmail.com

--

Last update: Mon 29 Dec 2020 AEDT