Face Anti-spoofing based on Convolutional Neural Networks

Abstract

Biometrics technologies have gained increasing attention across different sectors in the past decade. Face recognition has proven to be one of these successful biometric technologies. For example, law enforcement uses face recognition for faster investigations, banks for identity confirmation, and different organisations for access control. However, face recognition has shortcomings regardless of its high successes, just like any biometrics technology. Face recognition technology is still susceptible to face spoofing attacks despite great efforts made by different researchers to combat such attacks. The study proposes an anti-spoofing model based on deep learning methods. Three different pipelines are implemented based on convolutional neural network (CNN) architecture. A hyper tuned baseline CNN, a convolutional neural network based on AlexNet architecture, and a neural network based on VGG16 architecture. The study benchmarked pipelines using the available face anti-spoofing detection datasets - the NUAA and CelebA datasets. The study measures these performance metrics for all the pipelines: accuracy, precision, recall, F1 score, AUC, and Roc curve. All three pipelines provided good results when tested against the selected datasets.