Hand-Gesture-Based Biometric Verification and Identification Using Embedded-STQNet Deep Neural Architecture in Security-Oriented Systems

This study presents Embedded-STQNet, a dual-path deep neural architecture for surface electromyography (sEMG)-based biometric authentication. The network integrates a local-spectral CNN and a modified ResNet-18 encoder, with concatenated features passed through a shared 128-dimensional ℓ₂-normalized projection head. A three-stage metric learning strategy—Siamese contrastive pre-training, triplet-margin tuning, and quadruplet loss optimization—progressively enhances intra-class compactness and inter-class separation. We evaluate the model on a 28-channel sEMG spectrogram dataset collected across multiple trials from 43 subjects, performing 16 gestures in a multi-session setting. The Siamese configuration achieves a verification accuracy of 98.48% and the lowest Equal Error Rate (EER) of 0.0155, outperforming Triplet (0.0776) and Quadruplet (0.051) variants. In Rank-K identification, Siamese achieves 99.55% (Rank-1) accuracy in single-code and 99.52% in multi-code scenarios, while the Quadruplet model achieves 99.54% and 95.90%, and the Triplet model achieves 97.70% and 92.13%, respectively. Additional evaluations of DET, ROC, FAR/FRR, and clustering metrics confirm advanced discriminability, with a silhouette score of 0.88 and an intra/inter-class variance ratio of 0.0345. The model generates compact 128-D embeddings without relying on gesture classification, enabling generalizable and real-time suitability for identity recognition. Compared to prior EMG-based methods, Embedded STQNet offers improved accuracy, multi-day generalization, real-world viability for wearables, and real-time suitability for biometric systems.