Home // SECURWARE 2024, The Eighteenth International Conference on Emerging Security Information, Systems and Technologies // View article


Evaluating the Robustness of Kolmogorov-Arnold Networks against Noise and Adversarial Attacks

Authors:
Evgenii Ostanin
Nebojsa Djosic
Fatima Hussain
Salah Sharieh
Alexander Ferworn

Keywords: Kolmogorov-Arnold Network, KAN, MLP, FGSM, PGD, MNIST, Classification.

Abstract:
Abstract—Kolmogorov-Arnold Networks (KANs) is a new perspective direction in Machine Learning (ML) domain. KANs use spline functions to enhance interpretability and adaptability of the ML models. However, their robustness against Adversarial Attacks (AAs) has not been fully researched. This paper aims to address this gap by evaluating KAN performance under Gaussian noise and AAs, by using the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks. The objective of this paper is to assess the comparative robustness of KANs and Multi-Layer Perceptrons (MLPs) when exposed to Gaussian noise and adversarial attacks, aiming to identify areas of improvement for KANs and to provide insights into their performance under real-world, noisy conditions. The results show that KANs achieve higher accuracy than MLPs in a clean environment. At the same time, KANs demonstrate noticeable reduction in accuracy under conditions where increased noise and adversarial perturbations are present. KANs experience a more substantial accuracy drop under FGSM and PGD attacks compared to MLPs, which reveals critical areas for improvement and further research. The sensitivity of KANs to Gaussian noise further highlights their limitations in real-world scenarios. These findings underscore the need for further research to develop more resilient KAN architectures and better understand their role in secure ML systems.

Pages: 11 to 16

Copyright: Copyright (c) IARIA, 2024

Publication date: November 3, 2024

Published in: conference

ISSN: 2162-2116

ISBN: 978-1-68558-206-7

Location: Nice, France

Dates: from November 3, 2024 to November 7, 2024