Mobile Face Anti-spoofing with Deep Learning

Abstract

Asmore andmore applications are adopting facial recognition technologies, the underlying security issues are attracting increasing attention recently. Not only some simple attacks using photos or screen but also advanced devices like 3D mask and VR have posed threats to the current facial recognition system. Although there have been some proposed solutions to deal with such kind of problems, how to apply deep learning, computer vision and other related techniques to effectively and efficiently differentiate the real human faces from the fake ones is still an important field need to be investigated. This project aims at proposing a feasible and effective solution to prevent mobile face-spoofing with advanced techniques based on the current development of deep learning and computer vision. The objective is to detect the fake human faces with better accuracy compared to the currentmethods so that the systemcan reduce the present security risk. Due to the limitation of the computing resources onmobile devices, it is urgent to build a secure and accurate system which can differentiate real human faces and the fake ones. Therefore, this project is not only dealingwith finding a better algorithmto face anti-spoofing detection but also discovering the possibility to apply model compression to the existing models so that they could fit into the mobile devices. Meanwhile, the user experience like the speed of verification should not degrade. The first part of the project will be focused on the evaluation and improvement of the face anti-spoofing algorithms. Then, attention will be put on how to integrate the system to mobile devices. Through a series of testing over some popular mobile devices under various illumination situation, a robust and effective systemwill be built as the final product of this project.