Face spoofing indicator using deep learning

Abstract

Biometric face recognition technology is vital in security. With social media platforms such as Facebook, Instagram, YouTube, obtaining an individual’s photo or video is easy. With ill intentions, these imageries can be abused and exploited to attack face recognition-based biometric systems. This research provides an overview of presentation attacks (PA) and explores anti-spoofing techniques enabled through machine learning. I approached the issue as a binary classification problem and obtained over 40000 images of different ethnicities, separated into their respective classes of real and fake. I have also explored different techniques such as Visual Geometry Group (VGG)-esque architecture, transfer learning and eye blinking detection using state-of-the-art technologies such as TensorFlow, Keras, OpenCV, Scikit-learn. The algorithms are written in Python. Based on my findings, the results obtained were significant – a 99% accuracy in differentiating between spoofed and real faces.