Deep learning detection of facial biometric presentation attack

Face recognition systems have gained increasing importance in today’s society, which applications range from access controls to secure systems to electronic devices such as mobile phones and laptops. However, the security of face recognition systems is currently being threatened by the emergence of spoofing attacks that happens when someone tries to unauthorizedly bypass the biometric system by presenting a photo, 3-dimensional mask, or replay video of a legit user. The video attacks are perhaps one of the most frequent, cheapest, and simplest spoofing techniques to cheat face recognition systems. This research paper focuses on face liveness detection in video attacks, intending to determine if the provided input biometric samples came from a live face or spoof attack by extracting frames from the videos and classifying them by using the Resnet-50 deep learning algorithm. The majority voting mechanism is used as a decision fusion to derive a final verdict. The experiment was conducted on the spoof videos of the Replay-attack dataset. The results demonstrated that the optimal number of frames for video liveness detection is 3 with an accuracy of 96.93 %. This result is encouraging since the low number of frames requires minimal time for processing.