Cross-Architecture Evaluation of Deep Learning Models for Multimodal Biometric Verification with Score-Level Fusion
Comparing the deep learning architectures used for biometric verification in a fair manner has remained a difficult task. Researchers typically rely on different datasets, employ varying preprocessing steps, and adopt inconsistent evaluation criteria, all of which muddy the picture when trying to decide whether one network truly outperforms another. This work attempts to cut through that noise by fixing every variable except the backbone architecture itself. Residual Network-50) ResNet50), Efficient Network Version 2 Small (EfficientNetV2-S), and the Swin Transformer-Tiny (Swin-T) were trained and tested on the same LUTBIO and XJTU datasets, applying identical subject-disjoint splits across face, fingerprint, and palmprint modalities. The analysis measures Equal Error Rate (EER), Area Under the Curve, and True Acceptance Rate at False Acceptance Rate = 0.1%, and further examines five distinct score-level fusion strategies across every possible two- and three-modality pairing. The single-modality experiments reveal a clear pattern: Swin-T achieves a perfect 0.000% EER on faces, ResNet50 handles fingerprints best at 1.957% EER, and EfficientNetV2-S leads on palmprints with 0.140% EER. On LUTBIO, 10 out of 12 tested fusion configurations reach a reported EER of 0.000%. This study reports the effect sizes and selected confidence intervals for the key comparisons, allowing both their statistical and practical significance to be assessed. The results show that no single architecture is universally superior; the best-performing backbone is modality-dependent. Score-level fusion consistently improved verification accuracy over the best single modality, substantially reducing error rates. These findings imply that architecture choices for biometric systems should be modality-aware and that score-level fusion is an appropriate method for robust multimodal verification. The controlled protocol additionally provides a reproducible framework for fair architectural comparison in future biometric research.