Lightweight Vision Transformer-Based Multimodal Biometric Recognition Using Feature-Level Fusion of Iris and Finger Vein Patterns

Multimodal biometric systems have gained increasing attention due to their ability to improve the reliability and security of user authentication. This study presents FI-VeT (Feature Fusion-based Iris-Finger Vein Vision Transformer), a lightweight multimodal biometric framework that combines iris and finger vein modalities through feature-level fusion using dual ViT-Tiny encoders. The iris images are preprocessed using segmentation, CLAHE enhancement, morphological operations, and normalization, while the finger vein images undergo Mini-ROI extraction, CLAHE enhancement, and normalization to enhance discriminative vascular patterns. The extracted modality-specific representations are fused and classified using a compact multilayer perceptron architecture. The proposed model was evaluated using a multimodal dataset constructed from the CASIA-Iris-Thousand and FV-USM datasets under controlled acquisition conditions. Since no publicly available paired iris and finger vein dataset exists, the study aligned 492 iris classes with 492 finger vein classes to enable multimodal feature-level fusion and evaluation. Experimental results showed that FI-VeT achieved an accuracy of 99.66%, precision of 99.59%, recall of 99.54%, and an F1-score of 99.52%. The model further obtained a false acceptance rate (FAR) of 0.55%, false rejection rate (FRR) of 0.55%, and genuine acceptance rate (GAR) of 99.45%. Computational analysis indicated that the framework maintains low complexity with 12.23 million parameters, a model size of 46.65 MB, and approximately 4.5 GFLOPs during inference. In addition, the model achieved an inference latency of 10.59 ms and 94.4 FPS, supporting its suitability for real-time biometric authentication applications. The results suggest that lightweight transformer-based multimodal fusion can provide strong recognition performance while maintaining computational efficiency in controlled biometric environments.