The Approach To Secure Face Authentication Against Impostor Attacks

Over the past couple of decade use of control access based systems specifically face recognition based systems are the main source of verification and has been growing.Enhanced and much trustworthy face recognition system are coming into existence.Therefore,the authentication of user's information are dependent on different biometric traits like face,finger prints etc.But in spite of advance development in face recognition systems there is still space for improvement.Face biometric systems are vulnerable to impostor attacks like when someone tries to get access of biometric systems with fake identity or using counterfeit evidence which looks similar to real identity.A single 2D camera is continuously susceptible to spoofing attacks,vulnerability to biometric recognition systems is still an open research area.Which has left many unsolved challenges.Face spoofing attacks to biometric face recognition system is one of the issue left in this domain.Against user's identity recognition,different types of face potential attacks are main threats to the security of biometric systems.Fabricating fake biometrics traits are getting bad to worse,with more diverse and sophisticated face spoofing attacks are come into play.In this regards many hybrid,unimodal and multimodal biometric recognition methods have been proposed to secure the biometric systems.Different techniques incorporate different type of biometric schemes based on either software or hardware and in some scenario fusion of both.In this thesis,we design a novel method to safeguard the face biometric systems in adversarial environments from face spoofing attacks.Most of the previously proposed countermeasures either based on convolutional neural network(CNN)or hand-crafted techniques have some limitation.Hand-crafted based feature techniques has shown some progress but still not been able to achieve high accuracy because all these techniques were dependent on designing and selection of extracted features.Similarly,convolutional neural network(CNN)its self cannot learn the temporal information whereas most of face spoofing attacks are in the form of videos and information in time and space is critical in identifying spoofing attacks.Spatio-temporal feature learning helps to find the most distinguish clues between the binary classes either real attempt or impostor attacks.Our proposed method extract spatio-temporal information to differentiate between legitimate access and impostor videos or video sequences of photo attacks.Learning both spatio-temporal feature of videos we utilize both the hand-crafted and convolutional neural network.A supervised manner approach for learning spatial information using CNN followed by local binary pattern(LBP-TOP)to extract the temporal information.We cascade LBP-TOP with CNN to extract both space and time information into a single descriptor.To test the effectiveness of our method we further carried out extensive experiments on three of the most diverse face spoofing attacks datasets including the one locally created.For classification purpose we utilized the support vector machine(SVM)with non-linear RBF kernel which obtained the best result in the classification domain.