A stab at robustness through MadryLab's library.

The aim is to explore http://madry-lab.ml/ robustness package.

Dependencies

• Pytorch 3.7 or higher
• Cudatoolkit 11
• Scikit-learn
• Robustness library by Madry https://github.com/MadryLab/robustness

Datasets

• FashionMnist: This dataset is an interesting and an upgrade from the original MNIST used by Madry's team. The dataset can be cloned directly GitHub where the images are under data/fashion. git clone https://github.com/zalandoresearch/fashion-mnist.git
• CINIC: This is another interesting and an upgrade from the original CIFAR10 used by Madry's team. Download instructions are here
• COVID XRAY : This dataset can be downloaded from https://github.com/UCSD-AI4H/COVID-CT where the images are under the folder name : Images-processed as CT_COVID.zip and CT_NonCOVID.zip

Respository Structure The following repository is organized as such

• This readme file (markdown file)

• Two main folders: Robustness and Jupyter Notebooks. The former encompasses all the changes to Robustness library to make it work for the three datasets. those files have to be copied into the conda/python robustness library location. The second includes all the Notebook and python files needed to generate results specific to our project.

• Notebooks serve to generate PGD-perturbed examples, PGD-train and standard-train model. Additionally every file is named properly to assist end-users to run them as demos.

• Every dataset has the following demo/notebook(ipynb) files:

  • A standard-training notebook of a ResNet50 against the natural(original set) named: (DATASET)STD.ipynb
  • An adv-trained model using the Robustness library named: Using robustness to ADV train-DATASET.ipynb (please note that this file is an adaptation of this
  • A file to generate adversarial accuracy on a std-trained model named: (DATASET)STD-ADV validation.ipynb
  • A file to generate perturbed(PGD-attacked) examples named: (DATASET)Adv Examples pre-trained ADV.ipynb
  • A python file to generate a robust dataset(for both Fashion and CINIC) named: (Dataset)STD-RobustDS.py

• FashionMnist and Covid Xray required custom dataloaders. CINIC was already supported by PyTorch.