| Over the years, face anti-spoofing (FAS) has gained importance in various face recognition systems for acting as a robust security mechanism. However, its reliable deployment in authentication systems has been curtailed by its vulnerability to presentation attacks (PAs). With time, multiple types of PAs have emerged to render traditional methods of FAS unreliable. This paper proposes a novel Frame-Level Facial Anti-Spoofing Network (FL-FASNet) using an ensemble of three Convolutional Neural Networks (CNNs) for the detection of PAs. The three models are binary pixel-wise Xception (BPX) supervision CNN, Fully Convoluted Depth (FCD) CNN, and patch-based (PB) CNN. Further, this paper introduces a new training dataset called the Multi-SpoofDefend Dataset (MSDD) which contains a diverse range of images and PAs. Unlike existing methods, the proposed solution works only on the frame level making it light and suitable for deployment on small devices like smartphones with minimal computational and time overhead. Deepfake detection has also been incorporated into the methodology as they are a rising threat to security systems. The proposed system gives exceptional results of 3.8% Equal Error Rate (EER) and 3.48% Half Total Error Rate (HTER) which significantly outperforms existing models used for FAS. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.