Momentum Iterative Gradient Sign Method Outperforms PGD Attacks

Sreenivasan Mohandas, Naresh Manwani, Durga Dhulipudi

2022

Abstract

Adversarial examples are machine learning model inputs that an attacker has purposefully constructed to cause the model to make a mistake. A recent line of work focused on making adversarial training computationally efficient for deep learning models. Projected Gradient Descent (PGD) and Fast Gradient Sign Method (FGSM) are popular current techniques for generating adversarial examples efficiently. There is a tradeoff between these two in terms of robustness or training time. Among the adversarial defense techniques, adversarial training with the PGD is considered one of the most effective ways to achieve moderate adversarial robustness. However, PGD requires too much training time since it takes multiple iterations to generate perturbations. On the other hand, adversarial training with the FGSM takes much less training time since it takes one step to generate perturbations but fails to increase adversarial robustness. Our algorithm achieves better robustness to PGD adversarial training on CIFAR-10/100 datasets and is faster than PGD string adversarial training methods.

Download


Paper Citation


in Harvard Style

Mohandas S., Manwani N. and Dhulipudi D. (2022). Momentum Iterative Gradient Sign Method Outperforms PGD Attacks. In Proceedings of the 14th International Conference on Agents and Artificial Intelligence - Volume 3: ICAART, ISBN 978-989-758-547-0, pages 913-916. DOI: 10.5220/0010938400003116


in Bibtex Style

@conference{icaart22,
author={Sreenivasan Mohandas and Naresh Manwani and Durga Dhulipudi},
title={Momentum Iterative Gradient Sign Method Outperforms PGD Attacks},
booktitle={Proceedings of the 14th International Conference on Agents and Artificial Intelligence - Volume 3: ICAART,},
year={2022},
pages={913-916},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0010938400003116},
isbn={978-989-758-547-0},
}


in EndNote Style

TY - CONF

JO - Proceedings of the 14th International Conference on Agents and Artificial Intelligence - Volume 3: ICAART,
TI - Momentum Iterative Gradient Sign Method Outperforms PGD Attacks
SN - 978-989-758-547-0
AU - Mohandas S.
AU - Manwani N.
AU - Dhulipudi D.
PY - 2022
SP - 913
EP - 916
DO - 10.5220/0010938400003116