Signature Verification using Siamese Neural Network

Abstract

Handwritten signatures are a critical biometric for identity verification in numerous legal and financial contexts. However, the detection of falsified signatures is one of the most challenging tasks in document forensics, especially in the presence of skilled forgeries. To address the difficulty in classifying genuine signatures and skilled forgeries, as well as improving the current best systems with 7% verification error, this paper details a signature verification system utilizing a Siamese neural network. The system is a deep learning architecture built around two identical convolutional neural network (CNN) subnetworks with shared weights, designed to learn a discriminative similarity metric from pairs of signature images. It has been trained to learn a feature space where similar observations are placed in proximity by exposing the network to a pair of similar and dissimilar observations and minimizing the Euclidean distance between similar pairs while simultaneously maximizing it between dissimilar pairs. The model was evaluated on the CEDAR Signature Dataset using standard performance metrics. The proposed network achieved an accuracy of 100%, precision of 1.00, recall of 1.00, F1-score of 1.00, and an AUC-ROC of 1.00, demonstrating its potential for highly reliable and automated verification in real-world applications, including financial and administrative systems in Nepal.