Daftstone/Random-Directional-Attack

Random Directional Attack for Fooling Deep Neural Networks

Random Directional Attack for Fooling Deep Neural Networks

This project is for the paper "Random Directional Attack for Fooling Deep Neural Networks". Our implementation is based on cleverhans .

The code was developed on Python 3.6

1. Install dependencies.

Our experiment runs on GPU,, install this list:

pip install -r requirements_gpu.txt

2. Download dataset and pre-trained models.

Download dataset and extract it to the root of the program, which contains the MNIST, SVHN, and ImageNet-10 dataset.

Download pre-trained models and extract it to the root of the program.

3. Usage of python white_box.py

usage: python white_box.py [--data DATA_NAME] [--max_angle MAX_ANGLE]
               [--nb_dimensions NB_DIMENSIONS] [--is_train [IS_TRAIN]]
               [--eps EPS] [nb_epochs EPOCHS_NUMBER] [batch_size BATCH_SIZE]

optional arguments:
  --data DATA_NAME
                        Supported: MNIST, SVHN, CIFAR-10, ImageNet-10.
  --max_angle MAX_ANGLE
                        Maximum angle of rotation.
  --nb_dimensions NB_DIMENSIONS
                        Number of dimensions selected.
  --is_train [IS_TRAIN]
                        User this parameter to train online, otherwise remove the parameter.
  --eps EPS
                        The size of perturbations
  --nb_epochs EPOCHS_NUMBER
                        Number of epochs the classifier is trained.
  --batch_size BATCH_SIZE
                        Size of each batch of data

4. Example.

Use pre-trained model.

python white_box.py --data mnist --max_angle 180 --nb_dimensions 10 --eps 0.05

Train model online.

python white_box.py --data mnist --max_angle 180 --nb_dimensions 10 --eps 0.05 --is_train --batch_size 128

5. Usage of python black_box.py

usage: python black_box.py [--data DATA_NAME] [--max_angle MAX_ANGLE]
               [--nb_dimensions NB_DIMENSIONS] [--is_train [IS_TRAIN]] [--sub_is_train SUB_IS_TRAIN]
               [--eps EPS] [nb_epochs EPOCHS_NUMBER] [batch_size BATCH_SIZE]

optional arguments:
  --data DATA_NAME
                        Supported: MNIST, SVHN, CIFAR-10, ImageNet-10.
  --max_angle MAX_ANGLE
                        Maximum angle of rotation.
  --nb_dimensions NB_DIMENSIONS
                        Number of dimensions selected.
  --is_train [IS_TRAIN]
                        User this parameter to train target model online, otherwise remove the parameter.
  --sub_is_train [SUB_IS_TRAIN]
                        User this parameter to train subtitute model online, otherwise remove the parameter.
  --eps EPS
                        The size of perturbations
  --nb_epochs EPOCHS_NUMBER
                        Number of epochs the classifier is trained.
  --batch_size BATCH_SIZE
                        Size of each batch of data

6. Example.

Use pre-trained model.

python black_box.py --data mnist --max_angle 180 --nb_dimensions 10 --eps 0.05

Train model online.

python black_box.py --data mnist --max_angle 180 --nb_dimensions 10 --eps 0.05 --is_train --sub_is_train

7. Citation

If you want to use random directional atttack for attack in academic research, you are expected to cite

@article{luo2019random,
title={Random Directional Attack for Fooling Deep Neural Networks},
author={Luo, Wenjian and Wu, Chenwang and Zhou, Nan and Ni, Li},
journal={arXiv preprint arXiv:1908.02658},
year={2019}
}