A Study On Face Anti-Spoofing With Infrared Liveness Detection

In recent years,with the theoretical advances and practical commercial applications of face recognition technology,the society has been more and more in touch with and even relying on this new technology,in various aspects of daily life.However,due to the inherent vulnerability of leakage,fraud and attack of this biometric modality,the issue of face recognition security is also largely exposed,while providing great convivence.easy to be attacked,the problem of face recognition insecurity has gradually exploded.Therefore,the design of a face recognition system with anti-fraud and anti-attack ability(face anti-spoofing research)poses a crucial issue and has become an important research topic in information security of face recognition.The current research on face anti-spoofing has made some progresses,and many excellent methods or algorithms have appeared.However,these methods have either relatively low performance or even have lost their effectiveness as new attack methods evolve.Therefore,a study of newer methods for face anti-spoofing to deal with these new attack methods is quite necessary.This thesis combines the technique of infrared imaging and the theories of deep learning,and has made some investigative efforts on the topic of face anti-spoofing,from the following aspects:(1)To review current methods of face anti-spoofing,and to summarize the problems of current face anti-spoofing methods.This thesis reviews current methods of face anti-spoofing in the literature,and summarizes their common shortcomings: Firstly,the current works on face anti-spoofing are usually based on the use of photo and video replay attacks in visible light scenes,but the difference between real and fake faces in visible light is quite insignificant.Especially with the coming-up of high-resolution cameras,printers and electronic screens,as well as 3D printers,the fake face is generated so vividly that it almost looks the same as the real face,making it quite difficult to distinguish between the two;Secondly,the universal capabilities of face anti-spoofing methods are insufficient.Most anti-spoofing methods usually present good performance against a specific attack measure.But once the attack measure varies,the anti-spoofing method will dramatically lower in performance or even fail to work;Thirdly,the generalization ability of the face anti-spoofing algorithms is generally poor.the data collected at different times,by different cameras or under different conditions have large variance,which has led to the problem that a system trained under a database usually performs poor when test on a different database.(2)A new approach to face anti-spoofing is proposed based on the near infrared technique.This thesis proposes to use the near-infrared band and deep learning methods to study face anti-spoofing.Due to the great characteristics of near-infrared,there exists significant difference between the real face and the fake face collected by the near-infrared camera.In this thesis,the fake face data is collected based on the existing near-infrared data of the face,and the near-infrared face anti-spoofing database used in our experiment is built.We have achieved good results in the experiments on a single database,and also achieved good results in the cross-database experiments.Moreover,the convolutional neural network method used in deep learning is also superior to the two traditional methods listed in this article.(3)To study the feasibility of face anti-spoofing in the long-wave infrared band.After researching near-infrared face anti-spoofing,we further conducted face anti-spoofing research in the long-wave infrared band.Due to the distribution of blood vessels and blood circulation,the temperature distribution of the face is very rich.The real face collected by the long-wave infrared imager is significantly different from the photo face,video face,and 3D printed face,which can be easily distinguished.We conducted preliminary experiments on the anti-spoofing of long-wave infrared faces by using the attack method of placing the face behind the photo and pasting the face to the photo.We also explored the feasibility of anti-spoofing systems when using face silicone masks attack.Experiments have proved that face anti-spoofing based on the infrared band is feasible.The anti-spoofing performance of near-infrared is better than that that in the visible band,when tested against common fraud measures,and the generalization ability is greater;Experimental results of face anti-spoofing in the long-wave infrared band are in line with our expectations.When tested against common fraud measures,the discrimination ability can reach as high as 100%.However,other more realistic fraud measures,such as silica gel face mask and sophisticated face cosmetics,need further research.