Models will be published.

Training the models

Optional Arguments:
--attack is the attack the last model will be attacked with
--pgd_iters if attack is 1 (PGD) number of PGD iterations
--epsilon attack epsilon
--step_size attack step size
Perturbations:
If a .npy file will be specified in --pert_file it will train with it
If not, a seed (can also be specified with --seed or a random one is chosen) is used and they are sampled from a uniform distribution.

Training MNIST

Run the following to train the model:

python crp_train.py  --dataset 0 --batch_size 128 --pert_count 10 --pert_range 0.3 --epochs 50 --chkpt_path mnist_model.pth 

Training SVHN

Run the following to train the model:

python crp_train.py --dataset 2 --batch_size 128 --pert_count 10 --pert_range 0.0313 --epochs 350 --chkpt_path svhn_model.pth --seed 323

Training CIFAR-10

Run the following to train the model:

python crp_train.py  --dataset 1 --batch_size 128 --pert_count 10 --pert_range 0.06274 --epochs 700 --chkpt_path cifar10_model.pth

Evaluation

Instead of --chkpt_path use the argument --model_path. Use --attack 0 for FGSM and --attack 1 for PGD. For example, for CIFAR-10:

python crp_train.py --dataset 1 --batch_size 128 --pert_count 10 --pert_range 0.06274 --epochs 700 --model_path cifar10_model.pth --attack 1

Results

Evaluation of our best model against PGD (100 iterations) and FGSM

Dataset	FGSM	PGD
MNIST	98.89%	99.49%
SVHN 20pert	81.84%	80.21%
CIFAR-10	66.75%	51.63%