moorkh : Adversarial Attacks in Pytorch

moorkh is a Pytorch library for generating adversarial examples with full support for batches of images in all attacks.

About the name

The name moorkh is a Hindi word meaning Fool in English, that's what we are making to Neural networks by generating advesarial examples. Although we also do so for making them more robust.

Usage

Installation

• pip install moorkh or
• git clone https://github.com/akshay-gupta123/moorkh

import moorkh
norm_layer = moorkh.Normalize(mean,std)
model = nn.Sequential(
    norm_layer,
    model
)
model.eval()
attak = moorkh.FGSM(model)
adversarial_images = attack(images, labels)

Implemented Attacks

• EXPLAINING AND HARNESSING ADVERSARIAL EXAMPLES: FGSM
• ADVERSARIAL EXAMPLES IN THE PHYSICAL WORLD: IFGSM
• ON THE LIMITATION OF CONVULATIONSAL NEURAL NETWORK IN RECOGNIZING NEGATIVE IMAGES: Semantic
• ADDING NOISE: Noise
• TOWARDS DEEP LEARNING MODELS RESISTANT TO ADVERSARIAL ATTACKS: PGD\L2
• ESEMBLE ADVERSAIAL TRAINING: ATTACKS and DEFENSE: RFGSM

To-Do's

• Adding more Attacks
• Making Documentation
• Adding demo notebooks
• Adding Summaries of Implemented papers(for my own undestanding)

Contribution

This library is developed as a part of my learning, if you find any bug feel free to create a PR. All kind of contributions are always welcome!

References

• Adversarial=Robustness-Toolbox by IBM.
• Foolbox by Bethgelab.
• Cleverhans by Google brain
• Reliable and Interpretable Artificial Intelligence A Eth Zurich course
• Adversarial Robustness - Theory and Practice Tutorial by Zico Kolter and Aleksander Madry