Adversarial Training Library - Pytorch

Pytorch implementation of common adversarial training methods, including:

PGD-AT [1]
TRADES [2]
LBGAT [3]
MART [4]
STAT [5]
AT-AWP [6]
TRADES-AWP [6]
STAT-AWP [5]

Checkpoints can be downloaded here: 百度网盘(w5rr)

Training from Scratch

You can train from scratch as follows

python normal_train.py --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]
python at.py           --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]
python trades.py       --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]
python lbgat.py        --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]
python mart.py         --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]
python stat.py         --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]
python at_awp.py       --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]
python trades_awp.py   --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]
python stat_awp.py     --data [CIFAR10|CIFAR100] --arch [ResNet18|WideResNet]

Evaluate

You can perform a robustness evaluation on the pretrained weights:

python evaluate.py --data CIFAR10 --arch ResNet18 \
--checkpoint ./adversarial_training_checkpoints/stat_awp_resnet18_cifar10.tar

Results

All results are selected with the best robustness on the test set.

ResNet18 - CIFAR10

Method	Clean Acc	FGSM	PGD-20	PGD-100	CW-100
Natural	95.38	95.38	0.02	0.00	0.00
PGD-AT [1]	82.57	56.99	51.28	51.02	49.66
TRADES [2]	80.86	55.91	51.23	51.14	48.18
LBGAT [3]	76.40	56.69	53.19	52.98	50.41
MART [4]	78.59	58.54	54.87	54.58	49.14
STAT [5]	83.39	59.52	53.85	53.58	50.96
AT-AWP [6]	82.30	59.68	54.73	54.39	52.05
TRADES-AWP [6]	81.90	58.96	55.76	55.68	52.36
STAT-AWP [5]	83.03	60.20	56.35	56.25	52.64

ResNet18 - CIFAR100

Method	Clean Acc	FGSM	PGD-20	PGD-100	CW-100
Natural	78.70	8.30	0.00	0.00	0.00
PGD-AT [1]	56.76	31.96	28.91	28.89	26.89
TRADES [2]	54.77	30.02	27.90	27.70	24.33
LBGAT [3]	57.07	34.70	32.50	32.43	27.32
MART [4]	54.20	34.48	32.19	32.15	27.99
STAT [5]	57.86	32.97	30.55	30.46	26.47
AT-AWP [6]	58.75	35.47	32.57	32.50	30.00
TRADES-AWP [6]	59.06	34.13	31.77	31.76	27.62
STAT-AWP [5]	58.35	34.83	32.69	32.63	27.95

WideResNet-34-10 - CIFAR100

Method	Clean Acc	FGSM	PGD-20	PGD-100	CW-100
Natural	96.31	55.30	0.04	0.00	0.00
PGD-AT [1]	86.48	61.16	54.89	54.42	54.10
TRADES [2]	84.38	60.50	54.58	54.40	53.13
LBGAT [3]	79.47	60.56	56.74	56.58	54.08
MART [4]	83.01	61.65	56.72	56.36	53.15
STAT [5]	84.75	62.23	57.29	57.10	54.93
AT-AWP [6]	85.60	63.18	57.84	57.54	55.92
TRADES-AWP [6]	84.90	62.87	59.19	59.09	55.99
STAT-AWP [5]	86.41	64.83	60.31	60.03	56.58

Citation

If you are considering citing this work, please refer to the following:

@misc{yuan2023atp,
  title        = {Adversarial Training Library - Pytorch},
  author       = {Yuan, Xiaojian},
  year         = {2023},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/LetheSec/Adversarial_Training_Pytorch}},
}

Reference

[1] Towards Deep Learning Models Resistant to Adversarial Attacks

[2] Theoretically Principled Trade-off between Robustness and Accuracy

[3] Learnable Boundary Guided Adversarial Training

[4] Improving Adversarial Robustness Requires Revisiting Misclassified Examples

[5] Squeeze Training for Adversarial Robustness

[6] Adversarial Weight Perturbation Helps Robust Generalization