Research On Presentation Attack Detection Algorithm For Finger Vein Recognition

In the field of biometric recognition,finger vein recognition has become a research hot spot in recent years due to its characteristics of live identification,non-contact,and unsusceptible to external factors,and it has been applied to finance,education,and other fields.The vein feature information is the unique basis for the finger vein recognition systems to conduct user authentication.Once the user’s vein information is stolen,the attacker is likely to impersonate the legitimate user by forging the finger vein sample so that the recognition systems will suffer the threat of presentation attacks.In order to cope with such security problems,research on presentation attack detection for finger vein recognition systems is conducted,and the relevant work is as follows.(1)Due to the difficulty of finger vein image acquisition and falsification,the number of finger vein samples obtained is usually small.The effect of existing detection methods for finger vein presentation attacks is limited by sample size,which can lead to overfitting.Meanwhile,the pooling operation loses some valuable spatial information,resulting in poor recognition of finger offsets and axial rotation samples.Therefore,a new finger vein presentation attack detection method is proposed.This method introduces capsules that can store spatial info and applies Bayes’ theorem,which can optimize classification performance to the routing process between capsule layers.A new dataset for experiments that refers to existing attack methods is built in this thesis.The simulation results of this method display its good ability to distinguish between real and fake images of finger veins.(2)In the application scenarios of mobile devices and embedded devices,resource-constrained devices put forward higher requirements for the model complexity of the network,and it is urgent to strike a balance between the detection accuracy and the lightness of the model.In order to solve this problem,this thesis proposes a lightweight presentation attack detection method.In this method,the addition and displacement operations are applied instead of the multiplication operations to reduce the model overhead.Meanwhile,the central differential convolution module is constructed for feature extraction to improve the detection performance of presentation attacks.The simulation results on different datasets show that compared with other lightweight networks,the new model can ensure lower model complexity and higher authenticity detection accuracy.