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

KAN vs KAN: Examining Kolmogorov-Arnold Networks (KAN) Performance under Adversarial Attacks

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

Keywords: FGSM; MNIST; Kolmogorov-Arnold Networks; KAN; PGD; Classification.

Abstract:
Recent interest in applying Kolmogorov-Arnold Networks (KANs) to the Machine Learning (ML) domain has grown significantly. Different KAN implementations leverage various architectures, with the primary distinction being their use of different learnable activation functions. While recent studies have benchmarked and evaluated the performance of different KAN models, little attention has been given to their robustness against Adversarial Attacks (AAs). In our previous work, we compared the performance of a single KAN model to a Multi- Layer Perceptron (MLP) classifier under Gaussian noise and AAs, using the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks on the MNIST dataset. In this paper, we extend that analysis by comparing several popular KAN implementations subjected to the same attacks. We evaluate standard metrics, including accuracy, precision, recall, and F1- scores, using the MNIST dataset as in prior research. The aim is to empirically investigate how different activation functions influence the robustness of KAN models under AAs. Our results reveal substantial differences in accuracy loss across KAN models when exposed to AAs.

Pages: 17 to 22

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