Adversarial Training with pytorch

python main.py -a -v

Accuracy (WIP)

Model	Acc.
VGG16	--.--%
ResNet18	51.99%
ResNet50	--.--%
ResNet101	--.--%
MobileNetV2	--.--%
ResNeXt29(32x4d)	--.--%
ResNeXt29(2x64d)	--.--%
DenseNet121	--.--%
PreActResNet18	--.--%
DPN92	--.--%

Learning rate adjustment

I manually change the lr during training:

• 0.1 for epoch [0,50)
• 0.01 for epoch [50,60)

Resume the training with python main.py -r --lr=0.01 -a -v

References

1. Authors' code: MadryLab/cifar10_challenge

2. Baseline code: kuangliu/pytorch-cifar

Notes

To read more about Projected Gradient Descent (PGD) attack, you can read the following papers:

1. Towards Deep Learning Models Resistant to Adversarial Attacks

2. Adversarially Robust Generalization Requires More Data