Towards an Adversarially Robust Normalization Approach

This is code of the paper "Towards an adversarially Robust Normalization Approach".

Prerequisites

• PyTorch 0.4.0+
• Python 3.5+
• Cuda 8.0+

File structure

.
├── models                  # containt different neural networks
│    ├── create_model.py    # Select model
│    ├── norm.py            # Different normalizations are defineded in it
│    ├── resnet.py          # resnet for cifar is defined
│    ├── resnet_imagenet.py # resnet for imagenet, pytorch torch.vision resnet is modidfied
│    ├── vgg.py             # vgg for cifar 
├── utils                   # functions of many different types that are not related to main paper are defined here
│    ├── data.py            # functions for dataloder for diffeerent datasets 
├── attacks.py              # function to select adversarial attacks from advertorch package 
├── functions.py            # functions to use in training and testing 
├── test_cifar.py
├── test_imagenet.py
├── train_cifar.py
├── train_imagenet.py
├── run.sh
└── README.md

Training Examples

Training ResNet-50 on CIFAR-100 using RobustNorm

python train_cifar.py --dataset cifar100 --depth 20 --norm RNT --checkpoint checkpoints/cifar100-resnet50-RNT

Testing

python test_cifar.py --dataset cifar10 --model resnet --depth 20 --norm BN --resume checkpoints/Adverserial/cifar10-resnet20-BN/model_best.pth.tar --lr 0.1 --train-batch 128 --gpu 0

Train ResNet18 on ImageNet using RobustNorm

python3 train_imagenet.py --dataset imagenet --model resnet18 --norm RNT --gpu 0,1,2,3 --checkpoint checkpoints/imagenet/imagenet-resnet18-RN

References

Some parts of this code are based on this repository