Deep Ensemble Learning Approach for Face Anti-Spoofing Detection based on Pre-trained Models

In this paper, we propose a Weighted Deep Ensemble Learning (WDEL) to increase the overall accuracy of face anti-spoofing model by exploiting multiple learning architectures. Current anti-spoofing models based on one deep learning architecture are not able to extract all important information for discriminating real and fake faces. We therefore design a framework that fuses the information of five pre-trained deep learning models, Inception, Xception, VGG16, ResNet50 and MobileNet to create a high-quality feature representation against spoofing attacks. Moreover, we apply a weighted voting mechanism with each weight being reliant on the corresponding classifier's performance from real and spoof classification considering different metrics: precision for the non-spoof classification and recall for the spoof classification. The final prediction is obtained by aggregating these multiple classifiers using their optimal weights in an ensemble fashion. Additionally, we provide an experiment comparing the performance of these aforementioned pre-trained deep learning models. The comparison reveals that MobileNet, Xception, and ResNet50 exhibit the highest recall for spoofing and precision for real images, along with improved F1 score values in detection. This establishes them as optimal choices for facial spoofing detection.