(previously named as Boundary Attack++) Code for HopSkipJumpAttack: A Query-Efficient Decision-Based Adversarial Attack by Jianbo Chen, Michael I. Jordan, Martin J. Wainwright.
If you are using the code to compare HSJA with other algorithms, you will need to set your own counter for the number of model evaluations. "max_num_evals" in hsja.py is the maximum number of evaluations for estimating gradient (for each iteration), not the total number of model evaluations. To increase the total number of model evaluations, set a larger num_iterations.
The code for HopSkipJumpAttack runs with Python and requires Tensorflow of version 1.2.1 or higher. Please pip install the following packages:
numpytensorflowkerasscipy
We provide as an example the source code to run HopSkipJumpAttack on a ResNet trained on CIFAR-10. Run the following commands in shell:
###############################################
# Omit if already git cloned.
git clone https://github.com/Jianbo-Lab/HSJA
cd HSJA
###############################################
# Carry out L2 based untargeted attack on 5 samples.
python main.py --constraint l2 --attack_type untargeted --num_samples 5
# Carry out L2 based targeted attack on 5 samples.
python main.py --constraint l2 --attack_type targeted --num_samples 5
# Carry out Linf based untargeted attack on 5 samples.
python main.py --constraint linf --attack_type untargeted --num_samples 5
# Carry out Linf based targeted attack on 5 samples.
python main.py --constraint linf --attack_type targeted --num_samples 5
# Results are stored in cifar10resnet/figs/.
# For each image, the left-hand side is the original example and
# the right-hand side is its perturbed version.See main.py and hsja.py for details.
If you use this code for your research, please cite our paper:
@article{chen2019boundary,
title={HopSkipJumpAttack: A Query-Efficient Decision-Based Adversarial Attack},
author={Chen, Jianbo and Jordan, Michael I. and Wainwright, Martin J.},
journal={arXiv preprint arXiv:1904.02144},
year={2019}
}