Handwritten Signature Verification Using Explainable AI and Lightweight Hybrid CNN Model for Real-Time Applications

The rapid digitalization of financial and legal infrastructures has necessitated robust, automated methods for biometric authentication. Handwritten signatures remain the most socially accepted and legally binding method of identity verification. However, the proliferation of digital image editing tools has made signature forgery increasingly sophisticated, posing a severe threat to banking security and document integrity. While Deep Convolutional Neural Networks (DCNNs) have demonstrated remarkable success in offline signature verification, their deployment in real-world, resource-constrained environments remains a challenge due to high computational complexity and memory footprint. Furthermore, the “black box” nature of deep learning models limits their adoption in forensic and legal scenarios where interpretability is paramount. This thesis addresses these challenges by proposing a novel Lightweight Hybrid Convolutional Neural Network (LH-CNN) architecture specifically designed for offline signature verification on edge devices. The proposed model, LightX-SigNet, utilizes a Siamese framework integrated with Depth-wise Separable Convolutions and Residual Connections. This architectural choice drastically reduces the parameter count—by approximately 90% compared to standard VGG-based architectures—while maintaining high discriminative power. Crucially, this research integrates Explainable AI (XAI) using Grad-CAM++ to visualize the decision-making process. The XAI module generates heatmaps that highlight the specific regions of the signature (e.g., pen-pressure points, stroke discontinuities, curvature anomalies) that influence the verification decision, providing the transparency required for forensic admissibility. Extensive experiments were conducted on three benchmark datasets: CEDAR, GPDS- 160, and BHSig260-Hindi. The experimental results demonstrate that the proposed LightX-SigNet achieves an accuracy of 98.4% on the CEDAR dataset with a False Acceptance Rate (FAR) of 1.8%. Remarkably, the model achieves an inference time of 12ms on a standard mobile processor and occupies a disk space of less than 50KB after quantization. This work demonstrates that it is possible to achieve state-of-the-art accuracy in biometric verification suitable for real-time mobile applications without sacrificing the transparency required for legal auditability.