Deep learning approaches for online signature authentication: a comparative study of pre-trained CNN models

Abstract Authorization is essential for handling document assurance and security. Nowadays, it constitutes one of the top responsibilities for securing information and effectiveness in every domain. Technological advances have made interactions with machinery more effortless. As a result, the demand for authentication for various legitimate causes is growing rapidly. Therefore, biometric-based identification has dramatically accelerated. This is an improvement over the other approaches. The present work is intended to apply convolutional neural networks for mining features and supervised machine-learning techniques to verify handwritten signatures. Raw images of signatures were used to train the CNN models for feature extraction and data augmentation. In the present work, pre-trained CNN models, such as VGG16, Inception-v3, ResNet50, and Xception, were used to separate authentic from fake signatures. Supervised learning methods, including Logistic Regression and SVM, were used to classify features. The test data were obtained from the ICDAR 2011 Signature Dataset. The results obtained from the present work showed a clear improvement over traditional methods over 69 different signatures. VGG-16 with RMSProp achieved an impressive validation accuracy of 83%, demonstrating robustness with minimal overfitting. Compared with existing techniques, the proposed deep learning approach proved to be more accurate and reliable for signature verification.