Research

Fast Training of Provably Robust Neural Networks by SingleProp

AAAI

Authors

Published on

05/18/2021

Recent works have developed several methods of defending neural networks against adversarial attacks with certified guarantees. However, these techniques can be computationally costly due to the use of certification during training. We develop a new regularizer that is both more efficient than existing certified defenses, requiring only one additional forward propagation through a network, and can be used to train networks with similar certified accuracy. Through experiments on MNIST and CIFAR-10 we demonstrate improvements in training speed and comparable certified accuracy compared to state-of-the-art certified defenses.

Please cite our work using the BibTeX below.

@article{Boopathy_Weng_Liu_Chen_Zhang_Daniel_2021, 
title={Fast Training of Provably Robust Neural Networks by SingleProp}, volume={35}, 
url={https://ojs.aaai.org/index.php/AAAI/article/view/16840}, 
DOI={10.1609/aaai.v35i8.16840},
number={8}, 
journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
author={Boopathy, Akhilan and Weng, Lily and Liu, Sijia and Chen, Pin-Yu and Zhang, Gaoyuan and Daniel, Luca}, 
year={2021}, 
month={May}, 
pages={6803-6811} 
}
Close Modal