<![CDATA[Face recognition-based attendance systems can improve the convenience of attendance processes, but they remain vulnerable to presentation attacks, in which facial images displayed on another device are used to simulate a user's physical presence. At the Tangerang City Government, the Tangerang AYO application is used as the main attendance channel for civil servants (ASN) across government agencies (OPD), with attendance data linked to the calculation of Additional Employee Income (TPP). This study develops a face liveness detection model to distinguish between real and fake facial images using a CNN MobileNetV2 architecture with a transfer learning and two-phase fine-tuning approach. The dataset was constructed from actual ASN attendance data and consisted of real and fake images representing facial displays through a secondary device. The model was evaluated using a confusion matrix, accuracy, precision, recall, specificity, F1-score, ROC curve, and AUC. On the testing data (n=301), the model achieved an accuracy of 97.34%, precision of 96.13%, recall of 98.68%, specificity of 96.00%, F1-score of 97.39%, and AUC of 0.9934 at a threshold of 0.5. Grad-CAM analysis showed that the model produced different activation patterns between real and fake images, including contextual visual information in real images and screen-surface characteristics in fake images. A FastAPI-based REST API prototype achieved an average response time of 306.48 ms under warm conditions without a dedicated GPU, indicating the potential feasibility of the model as an additional verification layer for an ASN attendance system.]]>
Copyrights © 2026
