Courtesy based Deuce-Torrent Convolutional Networks for Face Spoofing Detection

Face spoofing attacks are a significant task in providing biometric security systems. This paper presents a new deep learning (DL) framework for face spoofing detection called Courtesy-based Deuce-Torrent Convolutional Networks (CDTCN). We adopt a two-stream convolutional strategy whereby the Deuce-Torrent module captures hierarchical spatio-temporal characteristics, and the Courtesy module improves the discriminability of spoofing articulation-specific features via adaptive feature pooling. The model presented here harnesses the power of multiple face anti-spoofing datasets to train a unified model for understanding what constitutes a genuine vs. spoofed face by utilizing texture, motion, and depth information. Comprehensive experimental results indicate that CDTCN achieves the best accuracy and generalization across datasets and is the most robust against various attack types, such as print, replay, and 3D mask attacks, over state-of-the-art methods. Results indicate the utility of our framework in practical face authentication settings and show that it is a promising choice for increasing biometric security.