Deep Learning - Based Handwritten Signature Verification System

The signature verification is the most important in biometric authentication, document validation and fraud prevention. Signature verification techniques from the previous generation are based on handcrafted features and statistical model, suffer from intra user variations and capable of forgeries made by skilled forgers. In this research we propose a deep learning based handwritten signature verification system that takes advantage of methodologies which have recently been developed in order to assure a level of accuracy, robustness and real time of signature verification. By integrating a Siamese Neural Network and a hybrid CNN - Transformer architecture, we learn to mimic both spatial as well as contextual dependencies in signatures. We also improve the model verification to be more accurate with the help of multi scale feature extraction, attention mechanisms and contrastive learning to distinguish the real ones from the forged ones. We further consider the use of GAN - based data augmentation to generate synthetic signatures that are more realistic and thus better lead to generalization of the model. A system that is designed to deployed on edge AI using TensorFlow Lite and both ONNX optimizations while being suitable for mobile and embedded devices. We also bring forth integration with blockchain for maintaining secure and tamper proof storage of verified signatures.