Contactless biometric verification from in-air signatures using deep siamese networks

In-air signature is a behavioral biometric trait that has gained increasing attention in recent years due to its contactless nature and potential for secure, hygienic, and remote authentication. Unlike traditional pen-and-paper or tablet-based systems, in-air signature methods capture signing gestures in three-dimensional space, typically using fingertip tracking or depth sensing, offering greater flexibility and accessibility in various application contexts. In this study, we developed a deep learning-based biometric verification model using in-air signature data collected from 25 participants. The collected dataset was structured into 200 positive (same-person) and negative (different-person) signature pairs, capturing both inter-person and intra-person variability. A Siamese Neural Network architecture based on Bidirectional LSTM layers and contrastive loss was used to learn a discriminative embedding space for signature verification. To rigorously evaluate generalization capability across users, we employed a customized cross-validation protocol based on the Leave Two Sample Out (LTSO) approach, a more stringent variation of the traditional Leave One Sample Out (LOSO) method, resulting in 300 unique train-test splits. The proposed system achieved strong overall performance, with an average accuracy of 85%, F1-score of 85%, and recall of 91%, indicating its effectiveness even with limited training data. These results demonstrate the feasibility of using in-air signatures as a practical, contactless biometric modality and support the viability of Siamese neural networks for learning person-specific patterns in motion-based verification tasks.