Hyperparameter Optimization of Lightweight Convolutional Neural Network Architectures Using Chicken Swarm Optimization for Improved Palm-Vein Biometric Performance

Abstract

ABSTRACT Palm vein recognition systems are increasingly vital for secure biometric authentication, requiring highly accurate and computationally efficient methods. Despite advances, achieving optimal performance remains challenging due to the high dimensionality and complexity of palm vein features. This study proposes a palm vein verification and identification system integrating lightweight convolutional neural networks (CNNs) with Chicken Swarm Optimization (CSO) for hyperparameter tuning, termed the CSO-CNN technique. Evaluated on a dataset of 400 images (150 genuine, 250 impostor), images were pre-processed for quality and region of interest extraction. The CSO-CNN optimized hyperparameters to extract discriminative features classified via a SoftMax layer. Results show that at an optimal threshold of 0.8, CSO-CNN achieved a False Acceptance Rate (FAR) of 6.00%, False Rejection Rate (FRR) of 12.00%, accuracy of 91.75%, and recognition time of 130.53 seconds. In contrast, the CNN recorded FAR of 7.20%, FRR of 14.67%, accuracy of 90.00%, and recognition time of 180.53 seconds. The Equal Error Rate (EER) further confirms CSO-CNN’s superiority with 10.67% versus 13.33% for CNN. Confusion matrix analysis revealed CSO-CNN correctly identified 132 genuine and rejected 235 impostors, misclassifying fewer samples. Paired t-tests confirmed significant improvements, with p-values of 0.006 and 0.000. CSO’s integration enabled efficient hyperparameter tuning, enhancing feature extraction and reducing recognition time. The palm vein’s larger surface area and richer features contributed to higher accuracy. These results demonstrate that CSO-CNN is a robust and efficient solution for palm vein biometrics, improving security and operational efficiency for practical deployment. Keywords: Palm vein recognition, vascular biometrics, lightweight convolutional neural networks, hyperparameter optimization, Chicken Swarm Optimization, deep learning, biometric authentication, metaheuristic algorithms, performance evaluation, recognition accuracy, resource constrained deployment.