Push-Pull layer for enhanced robustness of ConvNets [Paper]

Creating Python environment if not present (Optional)

## commands below creates a python virtual environment called  "myenv" and installs all libraries inside it.
python3 -m venv myenv
source ./myenv/bin/activate
pip install --upgrade pip
pip install jupyter ipython ipykernel magic-wormhole
ipython kernel install --user --name=myenv

pip3 install torch torchvision torchaudio
pip3 install tensorboard_logger tensorflow

Project Setup

# Clone the Repo
git clone https://github.com/robinnarsinghranabhat/Adversarial-Robustness-of-Push-Pull-CNN.git

# Goto to Project Root 
cd ./Adversarial-Robustness-of-Push-Pull-CNN

# Download the Corrupted-Cifar Dataset in above folder
wget https://zenodo.org/records/2535967/files/CIFAR-10-C.tar?download=1 -O CIFAR-10-C.tar

# Extract the Files inside the folder : `CIFAR-10-C`
tar -xvf CIFAR-10-C.tar

1. Without any Pushpull (Baseline Model)

Train for 120 epochs

Command below trains a BASELINE CNN Resnet-20 architecture. In example below, Batch size is 2040, epochs is 120.

After each iteration, output model and logs are saved at : ./experiments/resnet-cifar/resnet-20-no-pp-2048-120epoc Here, resnet-20-no-pp-2048-120epoc is name of experiment. And also folder name used to save results.

python -m train_direct -b 2048 --arch resnet --layers 20 --name resnet-20-no-pp-2048-120epoc  --print-freq 4 --epochs 120 --use-cuda

Test above model on CIFAR and Noisy-CIFAR

We use the same name used during training.

python -m test_corruption -b 2048 --arch resnet --name resnet-20-no-pp-2048-120epoc --layers 20 --corrupted-data-dir ./ --use-cuda --geom-transform

Test above model on FGSM distorted CIFAR and FGSM distorted NOISY-CIFAR

Specify parameter for : fgsm-epsilon

python -m test_corruption --fgsm-epsilon 0.01 -b 2048 --arch resnet --name resnet-20-no-pp-2048-120epoc --layers 20 --corrupted-data-dir ./ --use-cuda

Test above model on PGD distorted CIFAR and PGD distorted NOISY-CIFAR

Specify parameter for : pgd-epsilon

python -m test_corruption_pgd --pgd-epsilon 0.01 -b 2048 --arch resnet --name resnet-20-no-pp-2048-120epoc --layers 20 --corrupted-data-dir ./ --use-cuda --geom-transfom

2. With Pushpull at just first Convolution Layer

Train on CLEAN CIFAR

python -m train_direct -b 2048 --arch resnet --name resnet-20-pp-2048-120epoc --pushpull --layers 20 --print-freq 4 --epochs 120 --use-cuda

Test above model on CIFAR and Noisy-CIFAR

We use the same name used during training.

python -m test_corruption -b 2048 --arch resnet --name resnet-20-pp-2048-120epoc --pushpull --layers 20 --corrupted-data-dir ./ --use-cuda

Test above model on FGSM distorted CIFAR and FGSM distorted NOISY-CIFAR

Specify parameter for : fgsm-epsilon

python -m test_corruption  --fgsm-epsilon 0.01  -b 2048 --arch resnet --name resnet-20-pp-2048-120epoc --pushpull --layers 20 --corrupted-data-dir ./ --use-cuda

Test above model on PGD distorted CIFAR and PGD distorted NOISY-CIFAR

Specify parameter for : pgd-epsilon

python -m test_corruption_pgd  --pgd-epsilon 0.01  -b 2048 --arch resnet --name resnet-20-pp-2048-120epoc --pushpull --layers 20 --corrupted-data-dir ./ --use-cuda --geom-transform

3. With PushPull across all layers

Train

python -m train_direct --pushpull --pp-all -b 2048 --arch resnet --name resnet-20-all-pp-2048-120epoc --layers 20 --print-freq 4  --epochs 120 --use-cuda

Test on CIFAR and NOISY-CIFAR

python -m test_corruption --pushpull --pp-all -b 2048 --arch resnet --name resnet-20-all-pp-2048-120epoc --layers 20 --corrupted-data-dir ./ --use-cuda --geom-transform

Test above model on FGSM distorted CIFAR and FGSM distorted NOISY-CIFAR

Specify parameter for : fgsm-epsilon

python -m test_corruption --fgsm-epsilon 0.01 --pushpull --pp-all -b 2048 --arch resnet --name resnet-20-all-pp-2048-120epoc --layers 20 --corrupted-data-dir ./ --use-cuda

Test above model on PGD distorted CIFAR and PGD distorted NOISY-CIFAR

Specify parameter for : pgd-epsilon

python -m test_corruption_pgd --pgd-epsilon 0.01 --pushpull --pp-all -b 2048 --arch resnet --name resnet-20-all-pp-2048-120epoc --layers 20 --corrupted-data-dir ./ --use-cuda