Offline Signature Verification: Exploring Intra-Variability Across Time Intervals

Handwritten signatures remain a fundamental personal identifier, widely employed for authentication in many countries despite the advent of the digital age. Their inherent variability, influenced by factors, e.g., mood, writing speed, and writing tool, poses challenges for robust authentication systems. This paper focuses on analyzing the intra-variability of signatures collected intermittently over extended periods for individual signers. We propose a novel model utilizing an incremental graph convolutional network integrated with a reinforcement learning-based attention mechanism to capture these temporal variations effectively. For experimentation, we have created a database (say, SignIT) due to the unavailability of a dataset as per our requirement, comprising signatures from 100 individuals with 64 genuine and 64 forged samples each. Our experiments on the SignIT produced some interesting results. Additionally, we checked our model performance on BHSig-Hindi and BHSig-Bengali datasets to check the model efficacy on benchmark datasets, and obtained encouraging outcomes.