A lightweight CNN architecture integrating gradient and pore features for high-precision fingerprint spoof detection with visual explainability

Abstract

Fingerprint spoofing poses a persistent threat to the reliability of biometric authentication systems, particularly those employing low-cost sensors. This study aims to enhance the accuracy, efficiency, and interpretability of fingerprint presentation attack detection (PAD) using a lightweight and explainable deep learning approach. The proposed method introduces a novel convolutional neural network (CNN) architecture that incorporates gradient magnitude and pore-level feature maps alongside normalized grayscale images to form a three-channel input tensor. A MobileNet-based backbone is employed for feature extraction, further refined through a Convolutional Block Attention Module (CBAM) to emphasize spoof-relevant regions. Grad-CAM is integrated to provide visual interpretability of model predictions. The system is trained and tested on public PAD datasets including LivDet and MSU-FPAD, with evaluation metrics comprising accuracy, F1-score, AUC, EER, APCER, and BPCER. The proposed model achieves a classification accuracy of 98.0%, an F1-score of 0.98, and an AUC of 0.995. It demonstrates strong resilience against spoof attacks while preserving low inference latency, making it suitable for real-time edge deployment. The integration of gradient and pore-level biometric features within a lightweight CNN, coupled with attention-based refinement and visual explanation, significantly enhances spoof detection in fingerprint biometrics. The framework’s efficiency and interpretability position it as a viable solution for security-sensitive applications, such as digital forensics, mobile authentication, and access control in financial systems. Future extensions will target real-time deployment, multimodal fusion, and robustness against adversarial spoofs.