This repository is intended to be a well-groomed pytorch mirror of cleverhans. There's a rich literature on exploiting vulnerabilities in the robustness of neural nets, though the following resources make for a good introduction to the subject:
- Towards Deep Learning Models Resistant to Adversarial Attacks - Madry et. al
- Tricking Neural Networks: Create your own Adversarial Examples - ML@Berkeley Blog
- Intriguing properties of neural networks - Szegedy et. al
- Adversarial Examples: Attacks and Defenses for Deep Learning- Yuan et. al (late 2017 survey paper)
The remainder of this README will focus on how to set up mister_ed to build your own adversarial examples/defenses and offer a brief tour of the contents of the repository. The hope is that by leveraging mister_ed, you'll be able to sidestep much of the machinery used in adversarial examples and get straight to building new attacks and defenses.
This library uses Pytorch >=0.4 on python 2,3 to perform differentiation and other common computation required by machine learning models. Follow the link above to perform installation. It is also recommended that you leverage GPU's to speed up computation, though generating adversarial attacks (particularly on MNIST/CIFAR10) are typically less expensive than classical neural net training.
The best way to set up mister_ed right now is to clone this git repository.
git clone https://github.com/revbucket/mister_ed
and if you manage python packages with pip, the requirements can be installed via
pip install -r requirements.txt
Feel free to fork and play with the code.
To get started immediately, you'll need to ensure that you have access to a pretrained network, and a dataset. We'll use CIFAR10 as an example dataset. Configuration parameters are stored in mister_ed/config.json. The important parameters to set up right now are dataset_path and model_path.
dataset_path: if you already have datasets on your machine, simply set this to the directory where they live. Datasets can be downloaded using standardpytorch.torchvisionmethods.model_path: if you already have pretrained pytorch models on your machine, simply set this to the directory where they live. Pretrained models are saved as files ending in.th, using the standardtorch.save(model.state_dict(), ...)method.
To get you going as quickly as possible, run the
python scripts/setup_cifar.py
script to do the following:
- Ensure all dependencies are installed correctly
- Ensure that CIFAR data can be accessed locally
- Ensure that a functional classifier for CIFAR has been loaded. By default, pretrained CIFAR10 resnets from Yerlan Idelbayev are used.
Then there are tutorials_{1,2,3}.ipynb located in notebooks/ that contain an overview of this repository's contents and how to get started!