Research On Anti-spoofing In Face Recognition

Face recognition technology becomes mature and its commercial products are more and more widely used in our life. However, compared to fingerprint and iris, the face biometric characteristic is very easy to copy by photo and video, therefore, spoofing face is a big threat to the security of face recognition system. Live face detection technologies have been developed recently. However, the security of those technologies is not high and limited when used in practice applications. To achieve robust and high security live face detection system, live face detection are studied from physiological activities, context enviroment features and thermal infrared characteristic of live face in this paper. The main contributions of the work are as follows:1. The framework of liveness detection problem is discussed from biology, chemistry, physics and psychology. The important live characteristics of human face are introduced. A survey of anti-spoofing methods in face recognition system is given. The features, advantages and disadvantages of the publicly reported anti-spoofing methods are analyzed. The ideal characteristics of anti-spoofing method are also summarized here.2. Eye blink physiological behavior is studied and a blink-based liveness detection method is proposed. The feasibility of against photo spoofing by detecting blink is analyzed and it has the merits of non-intrusive, simply equipped and easily intergated. We formulate blink detection as inference in an undirected conditional graphical framework. A contiuous procedure of open, close and open states in blink is deeply descrbied by long-range dependencies. For purpose of quick and accurate recognition of the blink behavior, eye closity, an easily-computed discriminative measure is developed. This blink detecting method can work real time in low resolution video under uncontrolled light condition.3. Context enviroment of live face is studied and a context-based liveness detection method is proposed. We use scene context as a liveness clue outside live face. The feasibility and merits of scene context clue are analyzed, Scene context clue model is established. In the model, scene context clue extracting algorithm and matching algorithm is proposed. The difference between scene clues of live face and spoofing video can be detected by context clue matching score.4. Thermal infrared characteristic of live face is studied and a liveness detection method is proposed by correlating between thermal infrared and visible face images. We analyze how to make use of thermal infrared clue to detect live face. The correlation of IR-visible image pair is modelled and the influence of different face parts on correlation is also investigated. Final correlation score of our method is cumulated by weighted scores of different parts in face. It can greatly improve the ability to detect liveness.