Hidden Markovian Model Combined with Dynamic Mode Decomposition for Detecting Deception in Videos

Biometric authentication systems can be deceived in one way or other. Spoofing attacks to these biometric systems has greatly effected in verifying the identity of an individual. Detection of the spoofing attacks is a serious problem whereas face recognition systems and voice authentication systems are mostly vulnerable to spoofing. Traditional spoofing detection methods are not up to the mark due to the rapid increase in the hacking methodologies. Several spoofing attack detection methods have been proposed but each of them has its own drawbacks. The basic method used in detection of spoofed video uses prior knowledge regarding live face images such as eye blinking and lip movements since attack types are often unknown and very different from each other. Due to the lack of efficiency and accuracy whereas handling excessive attacks, specific cue which is peculiar to the attack must be developed. Thus a data driven approach with high accuracy and efficiency was developed. A combination of Dynamic Mode Decomposition, Principle Component Analysis and Hidden Markovian Model was introduced to provide a better security for the biometric authentication systems in videos. The project focus on both facial spoofing detection and voice spoofing detection in videos simultaneously. Principle Component Analysis and Dynamic Mode Decomposition together helps in the facial biometric authentication and Hidden Markov Model helps in the voice authentication. It considers various facial expressions, and even tilted angled faces in the spoofed samples. Local Binary Patterns, and Support Vector Machine were introduced for identifying print attacks, cut photo attacks and replay attacks. It results in better security, scalability, efficiency and accuracy.