VeinKAN: A Finger Vein Recognition Model Based on Kolmogorov–Arnold Networks

Abstract Finger vein recognition has become a secure biometric method known for its robustness against spoofing and environmental variations. Traditional methods, which often rely on Multi-layer Perceptrons (MLPs), face limitations in adaptability stemming from fixed activation functions and linear weight constraints. Kolmogorov–Arnold Networks (KANs) offer a novel architecture that enhances nonlinear learning capabilities to improve performance without significantly increasing computational overhead. This study proposes a KAN-based approach for finger vein recognition and evaluates its performance against established Convolutional Neural Network (CNN) models, including InceptionV3, EfficientNet, and MobileNetV3. Experiments on the FV_USM and SDUMLA-HMT benchmark datasets reveal that the proposed model achieves accuracies of 99.3 % and 96.2 %, respectively, surpassing conventional architectures. Despite a higher parameter count (34.81 million), the proposed model maintains an inference time of 1.0096 ms, which is comparable to InceptionV3 (1.006 ms) and notably faster than EfficientNet_B4 (1.349 ms). With a computational complexity of 539.12 MMAC, it supports the feasibility of biometric systems requiring high accuracy and efficient processing. These findings highlight KANs as a promising advancement in biometric recognition technologies.