Protocol-sensitive fingerprint embedding analysis using a lightweight hybrid CNN–Transformer framework

Fingerprint recognition systems are commonly evaluated using classification accuracy; however, high classification performance does not necessarily reflect the robustness or separability of the learned embedding space under different biometric acquisition protocols. This study investigates protocol-sensitive embedding behaviour using a lightweight hybrid CNN–Vision Transformer architecture, termed HyFiT, consisting of approximately 6.78 million parameters and a 128-dimensional L2-normalized embedding projection framework. Experiments were conducted using grayscale and colorized fingerprint representations across multiple evaluation protocols, including the standard multi-impression FVC2000 dataset, reduced-impression FVC2000(R) protocols, and the sparse many-class SOCOFing dataset. The proposed framework was evaluated using classification accuracy, verification-based metrics, embedding-distance statistics, receiver operating characteristic analysis, equal error rate, and low-dimensional embedding visualization. Experimental results demonstrated strong verification robustness and well-separated embedding distributions for FVC2000 protocols, achieving up to 92.58% classification accuracy with an AUC of 0.9736 and low equal error rates. In contrast, the SOCOFing protocol exhibited severe embedding overlap, degraded inter-class separability, and substantially reduced verification robustness despite comparatively stable classification behaviour. Furthermore, reduced-impression protocols revealed significant sensitivity of embedding geometry to impression scarcity and intra-class variability. These findings demonstrate that classification accuracy alone is insufficient for evaluating fingerprint representation quality and highlight the importance of verification-oriented, embedding-centric, and protocol-aware evaluation strategies for reliable biometric recognition systems.