Latent Space-Driven Knuckle-Dorsal Fusion: A CNN-Autoencoder Framework for Robust Biometric Authentication

Abstract

Biometric authentication has demonstrated its efficacy in securing access to authentic registered users. Physiological traits, particularly those that can be captured contactlessly, have garnered significant attention from the research community in recent years. One such trait is human hands, which offer a diverse range of feature sets from both the palmar and dorsal views. This paper proposes a novel framework for holistic matching of dorsal views of human hands acquired in a contactless setup. The knuckle and dorsal vein regions of human hands are examined for their individual recognition capabilities using deep features extracted through convolutional neural networks. Furthermore, the latent space representation of these vital biometric regions obtained from the autoencoder is investigated to assess the high matching performance with reduced dimensions of the features. Additionally, the combination of scores obtained from the individual matching performances is presented to explore the improvement achieved through various fusion strategies: knuckle-type, finger-type, all-knuckles (12 knuckle regions), and holistic (12 knuckle regions and the dorsal region). A comprehensive performance analysis is conducted using metrics such as equal error rate (EER), false acceptance rate, false rejection rate, receiver operating characteristics, and decidability index. The framework achieves the optimal EERs for both the original feature length and the latent space feature representation, reaching 4.91% and 4.32%, respectively.

Cite this article as: A. Mathur and R. Vyas, “Latent space-driven knuckle-dorsal fusion: A CNN-autoencoder framework for robust biometric authentication,” Electrica, 2026, 26, 0244, doi: 10.5152/electrica.2026.25244.