The evaluation code for maximally robust classifiers using optimization bias and the Fourier-Linf attack from the paper Bridging the Gap Between Adversarial Robustness and Optimization Bias , F. Faghri, D. J. Fleet, C. Vasconcelos, F. Pedregosa, N. Le Roux, 2021.
See requirements.txt
To train and evaluate a linear model using gradient descent and evaluate against Linf attack run:
python -m train --config.dim 100 --config.num_train 10\
--config.num_test 400 --config.optim.lr 0.1 --config.temperature 0.0001\
--config.adv.lr 0.1 --config.adv.norm_type linf --config.adv.eps_tot 0.5\
--config.adv.eps_iter 0.5 --config.adv.niters 100 --config.log_keys\
'("risk/train/loss","risk/train/zero_one","risk/test/zero_one","risk/train/adv/linf","weight/norm/l1")'\
--config.optim.name gd --config.optim.niters 1000 --config.model.arch linear\
--config.model.regularizer none --config.log_interval 1
To train a linear convolutional model and evaluate against Fourier-Linf attack run:
python -m train --config './config.py' --config.seed 2 --config.temperature\
0.0001 --config.num_test 1 --config.dim 100 --config.num_train 3\
--config.adv.norm_type dftinf --config.adv.lr 0.1 --config.adv.niters 1\
--config.adv.pre_normalize --config.adv.post_normalize\
--config.adv.eps_from_cvxpy --config.adv.step_dir dftinf_sd\
--config.log_interval 1000 --config.log_keys\
'("risk/train/zero_one","risk/train/adv/dftinf","weight/linear/norm/dft1","margin/dft1")'\
--config.enable_cvxpy --config.model.arch conv_linear_exp\
--config.model.nlayers 2 --config.model.regularizer none --config.optim.name gd\
--config.optim.lr 0.1 --config.optim.niters 10000\
--config.log_dir "runs/X"
To run the Fourier-Linf attack against CIFAR-10 models, clone the
AutoAttack repository and replace
autoattack/autoattack.py and autoattack/autopgd_pt.py with copies in the
autoattack directory in this repository and copy the new file
autoattack/norm.py to autoattack/. Modifications are based on the
AutoAttack commit
9b264b5.
Then to evaluate the Fourier-Linf robustness of a model (e.g. robust Linf
model of Wu2020Adversarial_extra) run:
mkdir -p runs/X && python eval_cifar10.py\
--model_name Wu2020Adversarial_extra --model_norm Linf --eps 8\
--path runs/X
The grid file grid/linear.py specifies hyperparameter configurations used in
experiments.
If you found this code useful, please cite the following paper:
@misc{faghri2021bridging,
title={Bridging the Gap Between Adversarial Robustness and Optimization Bias},
author={Fartash Faghri and Cristina Vasconcelos and David J. Fleet and Fabian Pedregosa and Nicolas Le Roux},
year={2021},
eprint={2102.08868},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
This code is based on the robust_optim project released under the Apache License 2.0.