Enhancement of Siamese Neural Network for Improved Signature Fraud Detection

This study enhances the Siamese Neural Network (SNN) in detecting signature fraud detection by addressing its critical challenges in feature extraction, difficulty handling class imbalances, and computational inefficiency. SMOTE was employed to balance the dataset, optimized training methodologies were applied, and the network architecture was redesigned to improve performance and scalability. Experimental evaluations were conducted using publicly available datasets, CEDAR and BHSig260, under a writer-independent setup, where the model was trained on one group of individuals and tested on unseen writers. The enhanced model demonstrated substantial improvements in performance metrics. The enhanced model achieved significant performance improvements, with accuracy rising from 67.61% to 99.65%, and F1-score from 0.0000 to 0.9944, with ROC-AUC from 0.5000 to 0.9989, The findings highlight the enhanced model's effectiveness in real-world applications, reinforcing public document security associated with signature forgery. This research contributes to the growing field of biometric verification, offering a scalable and adaptable solution tailored for the evolving demands of signature authentication.