Code for Parsimonious Black-Box Adversarial Attacks via Efficient Combinatorial Optimization

This code is for reproducing the results in the paper, Parsimonious Black-Box Adversarial Attacks via Efficient Combinatorial Optimization, accepted at ICML 2019.

Citing this work

@inproceedings{moonICML19,
    title= {Parsimonious Black-Box Adversarial Attacks via Efficient Combinatorial Optimization},
    author={Moon, Seungyong and An, Gaon and Song, Hyun Oh},
    booktitle = {International Conference on Machine Learning (ICML)},
    year={2019}
}

Installation

• Python 3.5
• TensorFlow 1.4.0 (with GPU support)
• opencv-python
• Pillow

Prerequisites

Cifar-10

1. Download Cifar-10 dataset from and decompress it.

wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
tar -xvzf cifar-10-python.tar.gz

2. Download an adversarially trained model from MadryLab and decompress it.

wget https://www.dropbox.com/s/g4b6ntrp8zrudbz/adv_trained.zip
unzip adv_trained.zip

3. Set DATA_DIR and MODEL_DIR in cifar10/main.py to the locations of the dataset and the model respectively.

ImageNet

1. Download ImageNet validation dataset (images and corresponding labels). Note that the validation images must be contained within a folder named val and the filename of validation labels must be val.txt.

• For images

mkdir val
wget http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_img_val.tar
tar -xf ILSVRC2012_img_val.tar -C val

• For labels

wget http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz
tar -xvzf caffe_ilsvrc12.tar.gz val.txt

2. Place the directory val and the file val.txt in the same directory.

3. Download a pretrained Inception-v3 model from Tensorflow model library and decompress it.

wget http://download.tensorflow.org/models/inception_v3_2016_08_28.tar.gz
tar -xvzf inception_v3_2016_08_28.tar.gz

4. Set IMAGENET_PATH in imagenet/main.py and MODEL_DIR in imagenet/tools/inception_v3_imagenet.py to the locations of the dataset and the model respectively.

How to run

• Cifar-10 untargeted attack

cd cifar10
python main.py --epsilon 8 --max_queries 20000

• ImageNet untargeted attack

cd imagenet
python main.py --epsilon 0.05 --max_queries 10000

• ImageNet targeted attack

cd imagenet
python main.py --targeted --epsilon 0.05 --max_queries 100000

Acknowledgements

This work was partially supported by Samsung Advanced Institute of Technology and Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01367, BabyMind).

License

MIT License