Research On Presentational Attack Detection Algorithms For Finger Vein Recognition Systems

Finger vein recognition is an emerging biometric technique for personal authentication that has garnered considerable attention and made great progress.Since finger vein is almost invisible to the naked eye under natural lighting conditions and can only be acquired using infrared illumination,it has a higher level of security than other biometrics techniques,such as face and fingerprint,and theoretically can effectively prevent attempted presentation attacks.But recent studies have shown that finger vein recognition systems are also vulnerable to presentation attacks from printed vein images.By reviewing the literature we found that several finger vein presentation attack detection(PAD)methods had been proposed,but none of them can effective solve this problem.In this thesis,we concentrated on the finger vein PAD and proposed 2 more effective methods including “finger vein presentation attack detection using total variation decomposition” and “finger vein presentation attack detection using convolutional neural network”,i.e.TV-LBP and FPNet.First,through our in-depth analysis,we found that more blurriness and noise tended to occur in forged vein images.Based on these different information,an effective and robust method,TV-LBP,was proposed.To avoid mutual interference between blurriness and noise information,total variation(TV)regularization was adopted to decompose original finger vein images into structure and noise components.Then a block local binary pattern(LBP)descriptor was utlized to encode blurriness and noise information on the structure and noise components respectively.At last we used a cascaded support vector machine model for classification,by which finger vein presentation attacks could be effectively detected.Second,considering the success of convolutional neural network(CNN)in feature learning for image classification and some biometric PAD tasks,such as face and fingerprint,we researched CNN for its possibility in finger vein PAD and designed a new specific shallow network,named FPNet,for this task.In the training phase,traning samples were extended tens or even hundreds of times by extracting large patches from original images.Then in the test phase,all patches' scores of the same finger vein image were integrated to combine global and local information and thus to make a higher quality decision.Finally,addition to the public-database “IDIAP FVD”,a new finger vein presentation attack database was constructed to evaluate fairly and reasonablely the performance of our approaches.Images of our intra-database were taken from 100 students that include index,middle and ring fingers of both hands.The shot number of each finger was 6,and thus 7200(3600 real and 3600 forged)images were collected in total.Extensive experimental results showed that,our two proposed methods were able to achieve an accuracy of 100% on the test sets of both databases,clearly outperforming state-of-the-art methods.