ItemBook Chapter

Classic Models, Modern Threats: A Study on Adversarial Attack and Defense for Traditional ML Models

Title
Classic Models, Modern Threats: A Study on Adversarial Attack and Defense for Traditional ML Models
Creator
Kalaiselvi, K.; Khundongbam, Alex; Steffyn, Kezya; Mangaiyarkarasi, T.
Description
Adversarial attacks are a serious threat to machine learning models, both for conventional architectures, like neural networks, and for more sophisticated frameworks, like Vision Transformers (ViTs). Although a lot of work has been done to defend state-of-the-art deep learning models against attacks like Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Gaussian noise perturbations, classical machine learning models like logistic regression, support vector machines (SVMs), and decision trees are relatively less explored despite their extensive use in situations where low computational complexity and high interpretability are needed. This work presents a rigorous evaluation of the adversarial vulnerability of binary and other classical models on the MNIST dataset and explores the effectiveness of various defense mechanisms, including adversarial training, input pre-processing (Gaussian smoothing), and defensive distillation. Experiments demonstrate that adversarial training is the most effective defense that improves model robustness with classification accuracies of up to 96% in all attack scenarios. In contrast, defensive distillation and input preprocessing make modest gains, with accuracy levels ranging from 61 to 81% based on the nature of the attack. Through adversarial threat analysis of typical machine learning models, this work points out their inherent susceptibility to adversarial perturbations and introduces robust defense techniques. These results identify the necessity for robust security and reaffirm the practical viability of typical models in the scenario of resource-constrained environments, contributing towards a more complete picture of adversarial defenses for the entire spectrum of machine learning architectures. The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
Source
Studies in Systems, Decision and Control;Volume;645;pp.241-258
Date
01-01-2026
Publisher
Springer Science and Business Media Deutschland GmbH
Subject
Adversarial attacks; Adversarial defense; Computational efficiency; Model Robustness
Coverage
Kalaiselvi K., Department of Computer Science, Kristu Jayanti University, Bengaluru, India; Khundongbam A., Department of Computer Science, Christ University, Karnataka, Bengaluru, India; Steffyn K., Department of Computer Science, Christ University, Karnataka, Bengaluru, India; Mangaiyarkarasi T., Department of Management, FOM-MBA SRMIST VDP Campus, Chennai, India
Rights
Restricted Access; Hardcopy may be available in the library
Relation
ISSN: 21984182;
Format
online
Language
English
Type
Book chapter

Citation

Kalaiselvi, K.; Khundongbam, Alex; Steffyn, Kezya; Mangaiyarkarasi, T., “Classic Models, Modern Threats: A Study on Adversarial Attack and Defense for Traditional ML Models,” CHRIST (Deemed To Be University) Institutional Repository, accessed June 17, 2026, https://archives.christuniversity.in/items/show/24091.

Output Formats