DEVELOPMENT OF AN AUTOMATED SIGNATURE VERIFICATION SYSTEM USING A SIAMESE NEURAL NETWORK (SNN)

Handwritten signature verification remains a widely used method of personal authentication in financial, legal, and administrative systems. However, manual verification methods are often prone to errors, inconsistency, and inefficiency, especially when dealing with large volumes of documents. This study presents the development of an automated signature verification system using a Siamese Neural Network (SNN), designed to distinguish between genuine and forged signatures based on learned similarity patterns. The system was developed using Python on Google Colab and implemented using deep learning techniques with TensorFlow and Keras. The dataset was sourced from Kaggle and pre-processed through grayscale conversion, resizing, normalization, and data augmentation to improve model generalization. The Siamese architecture consists of twin convolutional neural networks that extract feature embeddings from signature image pairs, which are then compared using Euclidean distance. The model was trained using contrastive loss to minimize the distance between genuine pairs and maximize the distance between forged pairs. Experimental results show that the proposed model achieved a training accuracy of 97.82% and a validation accuracy of 96.43%, while test performance recorded an accuracy of 96.11%, precision of 95.83%, recall of 94.44%, and an F1-score of 95.13%. The system also achieved a False Acceptance Rate of 2.22% and a False Rejection Rate of 5.56%, with an AUC of 0.987, indicating strong classification capability. The results demonstrate that the proposed system is reliable, efficient, and suitable for real-world signature verification applications in security-sensitive environments.