JS-SpoofNet: A jointly supervised parallel branched neural network for spoof detection

Abstract

Spoof detection in complex real-world conditions has always been challenging for the face anti-spoofing research community. Most existing datasets need more practical variations for spoof detection in the wild and thus generate the need for a more complex dataset encompassing the required diversities. The single-image-based anti-spoofing solutions for face recognition systems sometimes suffer from inconclusiveness in scenarios with cluttered backgrounds or variable illumination. To address these issues, we first collected a diverse spoof detection dataset, CSDiNE, with a wide range of illumination intensities and background variations. Secondly, we proposed a jointly supervised parallel branched neural network, JS-SpoofNet, that utilizes temporal cues derived from a video sequence for robust spoof detection. JS-SpoofNet consists of a parallel branched architecture with a spoof classification network aided by an auxiliary network that utilizes intermediate features from the main branch for depth estimation. We conducted an elaborate ablation study and performed an extensive performance evaluation of our proposed method. On the in-house CSDiNE dataset, the proposed spoof detector attained a minimum average classification error rate (ACER) of 0.94%. Furthermore, the designed neural network outperformed existing video-based state-of-the-art spoof detection methods on popular diverse benchmark datasets.