This is an implementation of the ReColorAdv adversarial attack and other attacks described in the NeurIPS 2019 paper "Functional Adversarial Attacks".
Clone this repository by running
git clone https://github.com/cassidylaidlaw/ReColorAdv
You can experiment with the ReColorAdv attack, by itself and combined with other attacks, in the getting_started.ipynb Jupyter notebook. You can also open the notebook in Google Colab via the badge below.
You can also install the ReColorAdv package with pip by running
pip install recoloradv
The script evaluate_cifar10.py will evaluate a model trained on CIFAR-10 against the adversarial attacks in Table 1 of the paper. For instance, to evaluate a CIFAR-10 model trained on delta (L-infinity) attacks against a ReColorAdv+delta attack, run
python recoloradv/examples/evaluate_cifar10.py --checkpoint pretrained_models/delta.resnet32.pt --attack recoloradv+delta
The script evaluate_imagenet.py will download a ResNet-50 trained on ImageNet and evaluate it against the ReColorAdv attack:
python recoloradv/examples/evaluate_imagenet.py --imagenet_path /path/to/ILSVRC2012 --batch_size 50
If you find this repository useful for your research, please cite our paper as follows:
@inproceedings{laidlaw2019functional,
title={Functional Adversarial Attacks},
author={Laidlaw, Cassidy and Feizi, Soheil},
booktitle={NeurIPS},
year={2019}
}
For questions about the paper or code, please contact claidlaw@umd.edu.