Dynamic oversampling-driven Kolmogorov–Arnold networks for credit card fraud detection: An ensemble approach to robust financial security

Abstract

Credit card fraud detection remains a persistent challenge in digital finance due to severe class imbalance, evolving fraud tactics, and the need for real-time analysis. Traditional detection systems often rely on static oversampling techniques and fixed feature sets, which limit their adaptability and robustness. This paper addresses these gaps by proposing a novel deep learning framework that combines Kolmogorov–Arnold Networks (KAN) with dynamic oversampling and ensemble feature selection. The dynamic oversampling strategy leverages both SMOTE and Generative Adversarial Networks (GANs) with variable sampling rates, reducing overfitting and enhancing generalization. Meanwhile, an ensemble feature selection mechanism integrates multiple metaheuristic algorithms to identify the most relevant features for fraud detection. The proposed approach, evaluated on three benchmark datasets, demonstrates strong improvements in adaptability over conventional deep learning models. This work offers a scalable, data-efficient solution for real-world fraud detection, improving resilience to data imbalance and evolving fraud patterns.