Dual-Angle Spatiotemporal Network (DAST-Net): A Hybrid LSTM-Transformer Architecture for Online Signature Verification in Human-Computer Interface Systems

Abstract

 Secure identity verification in human-computer interface (HCI) systems is critical for protecting information system access. Existing methods often fail to integrate fine-grained temporal dynamics with long-range structural dependencies and largely neglect sensor-derived pen kinematic data. This paper presents the Dual-Angle Spatiotemporal Network (DAST-Net), a hybrid architecture for signature verification. The framework incorporates azimuth and altitude angles, captured by standard HCI sensors, as novel dynamic cues. A stacked LSTM models short-term temporal patterns, while a Transformer encoder captures cross-stroke correlations. Evaluated on three datasets (self-constructed, SVC2004-Task2, MYCT-100), DAST-Net achieves 99.43% accuracy and a 0.89% equal error rate (EER), representing a relative reduction of 36–63% over state-of-the-art baselines. These results demonstrate that fusing dual-angle features within a hybrid LSTM-Transformer framework provides a robust and implementable solution for enhancing information system security in HCI applications.