Research on Deep Learning-Based Authentication Methods for E-Signature Verification in Financial Documents

Abstract

This paper presents a novel deep learning-based authentication method for electronic signature verification in financial documents. The proposed system introduces a comprehensive framework integrating YOLOv5-based signature detection, advanced preprocessing techniques, and a Siamese neural network architecture for verification. The system employs a customized feature extraction network incorporating residual connections and attention mechanisms to capture local and global signature characteristics. The implementation includes adaptive preprocessing pipelines and sophisticated loss functions optimized for signature verification tasks. Experimental evaluation on a dataset of 25,000 signature samples from 500 individuals demonstrates superior performance, achieving 98.5% accuracy in verification tasks with a false acceptance rate of 1.2% and a false rejection rate of 1.5%. The system maintains robust performance across various document conditions, demonstrating only a 4.2% accuracy reduction under poor resolution scenarios. Security analysis validates system resilience against adversarial attacks, achieving a 96.5% detection rate. The comprehensive evaluation demonstrates significant improvements over existing accuracy and computational efficiency methods, establishing new benchmarks for signature verification in financial applications. The proposed methodology addresses critical challenges in financial document security while maintaining practical applicability in real-world environments.