JoWUA

Abstract

Face spoofing poses a serious threat to biometric systems. Although deep learning models achieve high accuracy in detecting spoofed faces, their decision-making process often lacks transparency, and their decisions are often difficult to understand. This research presents a region-wise explainability approach to enhance transparency and interpretability in face spoofing detection. The method focuses on spoof-prone facial regions, such as the eyes, nose, and mouth, identified using landmark-based segmentation. A ResNet18 model integrated with a Convolutional Block Attention Module (CBAM) is trained separately on full-face and cropped facial regions. The model achieves high accuracy across all inputs, with the full-face model attaining 99.22% test accuracy, while the nose, eyes, and mouth regions also show strong performance with accuracies above 97%. In addition, all regions achieve a zero false acceptance rate, which is important for secure biometric systems. Subsequently, Gradient-weighted Class Activation Mapping, known as Grad-CAM, is applied across all face regions to visualize the regions that influence the model’s decisions. The generated heatmaps show that the full face yields the highest prediction confidence of 0.956 among the other regions. Further, a confidence drop analysis is performed by masking the most important regions identified by Grad-CAM. This helps to understand how much each facial region contributes to the final decision and verifies whether the model is focusing on meaningful spoof-related features.