Enhancing Face Spoofing Attack Detection: Performance Evaluation of a VGG-19 CNN Model

Abstract

With the wide use of facial verification and authentication systems, the performance evaluation of Spoofing Attack Detection (SAD) module in the systems is important, because poor performance leads to successful face spoofing attacks. Previous studies on face SAD used a pretrained Visual Geometry Group (VGG) -16 architecture to extract feature maps from face images using the convolutional layers, and trained a face SAD model to classify real and fake face images, obtaining poor performance for unseen face images. Therefore, this study aimed to evaluate the performance of VGG-19 face SAD model. Experimental approach was used to build the model. VGG-19 algorithm was used to extract Red Green Blue (RGB) and deep neural network features from the face datasets. Evaluation results showed that the performance of the VGG-19 face SAD model improved by 6% compared with the state-of-the-art approaches, with the lowest equal error rate (EER) of 0.4%. In addition, the model had strong generalization ability in top-1 accuracy, threshold operation, quality test, fake face test, equal error rate, and overall test standard evaluation metrics.