Improving the resilience of complex neural networks to attacks using the fast gradient method with adversarial training
DOI:
https://doi.org/10.21303/2585-6847.2025.004076Keywords:
deep learning, adversarial attacks, fast gradient sign method, adversarial learningAbstract
Deep neural networks used in finance, cybersecurity, and critical infrastructure are vulnerable to hostile perturbations, where small, targeted changes to the input data lead to large classification errors. In this paper, let’s evaluate a convolutional neural network on the Modified National Institute of Standards and Technology (MNIST) dataset under Fast Gradient Signature Method (FGSM) attacks to test whether one-step adversarial learning can restore reliability without costly multi-step defenses. The goal is to limit the drop in accuracy from 99.05% on clean data under FGSM to 0.1 while maintaining nominal performance. The approach augments each training batch with FGSM samples in a 1:1 ratio and applies a one-step minimax optimization that jointly optimizes accuracy on clean and noisy data. The resulting model achieves an accuracy of 98.94% on clean images and 99.38% on noisy images, gaining 13.89 percentage points of attack resistance at the expense of only 0.11 percentage points on clean data. This is due to the broader, more stable internal representations that generalize across clean and noisy variants, rather than retraining a narrow, gradient-sensitive decision boundary. FGSM's single-stage training achieves reliability close to multi-stage schemes such as Projected Gradient Descent (PGD), while requiring nearly 10 times less computation and no additional architectural complexity, which is suitable for resource-constrained applications including digit recognition, Optical Character Recognition (OCR), fully automated public computerized Turing test (CAPTCHA), biometric authentication, and intrusion detection, although deployments in high-risk environments still require testing for resilience to stronger attacks and more complex datasets
Downloads
References
Kwon, H. (2021). MedicalGuard: U-Net Model Robust against Adversarially Perturbed Images. Security and Communication Networks, 2021, 1–8. https://doi.org/10.1155/2021/5595026
Waghela, H., Sen, J., Rakshit, S. (2024). Robust Image Classification: Defensive Strategies against FGSM and PGD Adversarial Attacks. 2024 Asian Conference on Intelligent Technologies (ACOIT), 1–7. https://doi.org/10.1109/acoit62457.2024.10941671
Chowdhury, K. (2024). Adversarial Machine Learning: Attacking and Safeguarding Image Datasets. 2024 4th International Conference on Ubiquitous Computing and Intelligent Information Systems (ICUIS), 1360–1365. https://doi.org/10.1109/icuis64676.2024.10866337
Wong, E., Rice, L., Kolter, J. Z. (2020). Fast is better than free: Revisiting adversarial training. arXiv. https://doi.org/10.48550/arXiv.2001.03994
Yang, B., Xu, K., Wang, H., Zhang, H. (2022). Random Transformation of image brightness for adversarial attack. Journal of Intelligent & Fuzzy Systems, 42 (3), 1693–1704. https://doi.org/10.3233/jifs-211157
Chang, T.-J., He, Y., Li, P. (2018). Efficient Two-Step Adversarial Defense for Deep Neural Networks. arXiv. https://doi.org/10.48550/arXiv.1810.03739
Zhang, Q., Zhu, X., Song, J., Gao, L., Shen, H. T. (2022). Staircase Sign Method for Boosting Adversarial Attacks. arXiv. https://doi.org/10.48550/arXiv.2104.09722
Gupta, V., Narayan, A. (2023). Do we need entire training data for adversarial training? arXiv. https://doi.org/10.48550/arXiv.2303.06241
Jia, X., Chen, Y., Mao, X., Duan, R., Gu, J., Zhang, R. et al. (2024). Revisiting and Exploring Efficient Fast Adversarial Training via LAW: Lipschitz Regularization and Auto Weight Averaging. IEEE Transactions on Information Forensics and Security, 19, 8125–8139. https://doi.org/10.1109/tifs.2024.3420128
Lupart, S., Clinchant, S. (2023). A Study on FGSM Adversarial Training for Neural Retrieval. Advances in Information Retrieval, 484–492. https://doi.org/10.1007/978-3-031-28238-6_39
Li, Z., Liu, L., Wang, Z., Zhou, Y., Xie, C. (2022). Bag of Tricks for FGSM Adversarial Training. arXiv. https://doi.org/10.48550/arXiv.2209.02684
Jia, X., Zhang, Y., Wu, B., Wang, J., Cao, X. (2022). Boosting Fast Adversarial Training With Learnable Adversarial Initialization. IEEE Transactions on Image Processing, 31, 4417–4430. https://doi.org/10.1109/tip.2022.3184255
Downloads
Published
Versions
- 2025-11-28 (Version of Record 1.0)
Issue
Section
License
Our journal abides by the Creative Commons CC BY copyright rights and permissions for open access journals.
Authors, who are published in this journal, agree to the following conditions:
1. The authors reserve the right to authorship of the work and pass the first publication right of this work to the journal under the terms of a Creative Commons CC BY, which allows others to freely distribute the published research with the obligatory reference to the authors of the original work and the first publication of the work in this journal.
2. The authors have the right to conclude separate supplement agreements that relate to non-exclusive work distribution in the form in which it has been published by the journal (for example, to upload the work to the online storage of the journal or publish it as part of a monograph), provided that the reference to the first publication of the work in this journal is included.
