Facial-Liveliness-Verification for Monocular Real-Time-Systems

Abstract

Face-recognition is becoming the go-to authentication method. It is convenient: simply look at the camera for instant recognition. Attackers, however, can expose vulnerabilities by “replaying” an enrolled user. The primary concern here is the physical-spoof-attack. Attackers can acquire a representative image from social media and create a realistic looking facsimile (e.g., paper-mask) for authentication. This attack is rather popular for its efficacy and simplicity; despite this, there are few reliable monocular detection methods. Alternatively, attackers can tamper the camera stream by placing an injection device. The face-swap-attack similarly presents an acquired image of the victim, this time as a photo-realistic image alteration using machine-learning. This attack is new and does not yet have a computationally efficient means of detection. The goal of this dissertation is to address both problems in a fashion that is monocular, single-frame and computationally efficient. A series of four physics-informed facial-liveliness-verification frameworks are presented to achieve these goals. Performance evaluation shows best-in-class accuracy where all algorithms are optimized for real-time-systems. These results are discussed and concluded with proposed future works.