Official Code for AdvRush: Searching for Adversarially Robust Neural Architectures (ICCV '21)
Python == 3.6.12, PyTorch == 1.2.0, torchvision == 0.4.0
cd advrush && python train_search.py --batch_size 32 --gpu 0 --epochs 60 --a_gamma 0.01 --a_warmup_epochs 50 --w_warmup_epochs 60 --loss_hessian loss_cure
cd advrush && python adv_train.py --batch_size 64 --gpu 0 --epochs 200 --adv_loss pgd --arch ADVRUSH
Prior to the evaluation process, add all necessary checkpoint files (preferably in the form of .pth.tar) to the /eval/checkpoints folder. To conduct white-box attacks,
cd eval &&
python pgd_attack.py --white-box-attack True --test-batch-size 10 --arch [arch_name] --checkpoint [./checkpoints/file_name.pth.tar] --data_type [cifar10/svhn]
To conduct black-box attacks,
cd eval &&
python pgd_attack.py --test-batch-size 10 --target_arch [target_arch] --target_checkpoint [./checkpoints/target_file.pth.tar] --source_arch [source_arch] --source_checkpoint [./checkpoints/source_file.pth.tar] --data_type cifar10
DARTS: Differentiable Architecture Search [ICLR '19] code paper
Robustness via Curvature Regularization, and Vice Versa [CVPR '19] code paper
Tradeoff-inspired Adversarial Defense via Surrogate-loss Minimization [ICML '19] code paper