dongji1111 / featurescatter

Feature Scattering Adversarial Training

Feature Scattering Adversarial Training (NeurIPS 2019)

Introduction

This is the implementation of the "Feature-Scattering Adversarial Training", which is a training method for improving model robustness against adversarial attacks. It advocates the usage of an unsupervised feature-scattering procedure for adversarial perturbation generation, which is effective for overcoming label leaking and improving model robustness. More information can be found on the project page: https://sites.google.com/site/hczhang1/projects/feature_scattering

Usage

Installation

The training environment (PyTorch and dependencies) can be installed as follows:

git clone https://github.com/Haichao-Zhang/FeatureScatter.git
cd FeatureScatter
python setup.py install

Tested under Python 3.5.2 and PyTorch 1.2.0.

Train

Specify the path for saving the trained models in fs_train.sh, and then run

sh ./fs_train.sh

Evaluate

Specify the path to the trained models to be evaluated in fs_eval.sh and then run

sh ./fs_eval.sh

Reference Model

A reference model trained on CIFAR10 is here.

Cite

If you find this work is useful, please cite the following:

@inproceedings{feature_scatter,
    author = {Haichao Zhang and Jianyu Wang},
    title  = {Defense Against Adversarial Attacks Using Feature Scattering-based Adversarial Training},
    booktitle = {Advances in Neural Information Processing Systems},
    year = {2019}
}

Contact

For questions related to feature-scattering, please send me an email: [email protected]