A Comparative Study of CNNs and Siamese Networks for High Accuracy Arabic Handwritten Signature Verification رسالة ماجستير

Abstract

In the realm of legal, financial, and governmental systems, handwritten signature verification is remains a significant biometric authentication method. Although deep learning models reach near-perfect performance on Latin-script signatures such as those provided in the CEDAR dataset, performance has been limited by high intra-writer variability, cursive nature, and diacritical complexity in Arabic signatures. This research fills this void with a rigorous, script-specific evaluation framework. We then created a new standardized Arabic signature dataset, comprising 55 participants (24 genuine and 24 skilled forgeries per writer), mimicking the CEDAR's framework for multi-script comparison. Importantly, Siamese networks were trained using a 45/10 writer-independent protocol — a statistically sound model which produced transformative results: The VGG16 Siamese model produced 99.29% accuracy and 0.9930 F1-score and 0.9997 AUC on Arabic signatures, matching its 100% performance on CEDAR. This contradicts the popular belief that Siamese architectures are irreparably ill-suited for Arabic script, showing instead that the evaluation protocol design is the focus. On the other hand, CNNs (e.g., MobileNetV2 94.92% accuracy) were tested in a sample-level split and although their effectiveness is high, they do not match up well with real identity-conditioned verification. All Siamese models trained on Google Colab’s T4 GPU over 11–29 minutes which accounts for a more than 90% reduction from previous literature, enabling real-time deployment. These results demonstrate that Arabic signature verification is not inherently more difficult but rather requires suitable protocols, suitable architectures, optimized pipelines. We provide a foundation for reliable secure and practical implementation of accurate biometric systems in Arabic-native populations.