Modeling Dynamic Signatures through Two-Link Robotic Arm Biomechanics with Torque Features and CNN–BiLSTM for Forgery Detection

Abstract

Dynamic signature verification remains a challenging biometric problem due to the subtle neuromotor variations that differentiate genuine handwriting from skilled forgeries. This work introduces a biomechanics-inspired approach that models the human signing process as a two-link robotic arm, enabling the extraction of joint-space kinematic and torque-driven dynamic features from standard pen-trajectory data. Raw signature coordinates from the SVC2004 Task-2 dataset are first normalized and transformed into joint angles using inverse kinematics. Angular velocities, angular accelerations, and torque estimates are then derived through numerical differentiation, creating a six-dimensional temporal representation that captures the underlying neuromotor effort exerted during writing. These torque-enhanced sequences are fed into a hybrid deep learning framework combining one-dimensional CNN layers for local pattern extraction with a bidirectional LSTM network for temporal dependency modelling. The proposed system is trained using an 80/20 stratified split and evaluated using classification and verification metrics. Experimental results demonstrate strong discrimination between genuine signatures and skilled forgeries, confirming that torque-based biomechanical cues encode writer-specific motion dynamics not evident in spatial trajectory features alone. This study establishes two-link robotic arm biomechanics as an effective modelling paradigm for dynamic signatures and highlights torque-driven features as a promising direction for next-generation biometric authentication systems.

Keywords