hirokiadachi / customized-adversarial-training

Customized-Adversarial-Training

Pytorch implementation of "CAT: Customized Adversarial Training for Improved Robustness".
Authors: Minhao Cheng, Qi Lei, Pin-Yu Chen, Inderjit Dhillon, and Cho-Jui Hsieh
Paper: https://arxiv.org/abs/2002.06789

Training details

Model: WideResNet34-10
Epochs: 200
Batch size: 128
Optimizer: SGD
(momentum: 0.9)
(initial learning rate: 0.1)
(80th: 0.01)
(140th: 0.001)
(180th: 0.0001)

• CIFAR10

python train.py --dataset cifar10 --num_classes 10 --loss_type xent --gpus 0
python train.py --dataset cifar10 --num_classes 10 --loss_type mix --gpus 0

• CIFAR100

python train.py --dataset cifar100 --num_classes 100 --loss_type xent --gpus 0
python train.py --dataset cifar100 --num_classes 100 --loss_type mix --gpus 0

	Natural	FGSM	PGD-10	PGD-20	PGD-100	PGD-100(CW loss)	CW-20 (l2)	APGD-CE	APGD-DLR
CIFAR-10, Xent	93.73	78.86	71.53	68.12	64.36	64.83	68.71	56.12	23.91
CIFAR-10, Mix	93.66	80.57	73.80	68.61	62.38	62.50	73.30	53.85	26.16
CIFAR-100, Xent	72.14	46.64	40.79	39.42	37.80	20.17	26.04	34.80	13.17
CIFAR-100, Mix	72.32	48.29	42.46	40.72	39.09	19.92	27.21	36.44	13.15

customized-adversarial-training's Issues

Slow Training Progress

Dear hirokiadachi-san,

Thank you for your work and for sharing this implementation. I am now trying to reproduce the results, but 41 epochs into the training, I am still only getting a ~20% robust accuracy (natural accuracy is ~87%). I am hoping to ask if this is expected. Do you remember seeing the robust accuracy increase after more epochs (i.e., after reducing the learning rate at epoch 80), or is there likely a bug in my code (I made some minor modifications for compatibility with the newest PyTorch)? Really appreciate your help!

Sincerely,
Yatong