Adaptive Face Anti-Spoofing Algorithm Based On Multimodal Feature Fusion

In this era,Internet of things and information security technology are developing rapidly,equipment security and information security have attracted more and more attention.The face information encryption algorithm come into being.The main purpose of face fraud attack is to impersonate the identity of a specific person.The application scenarios are mainly customs clearance,replacing attendance and stealing accounts for payment.Common fraud methods mainly include printing attack(i.e.cheating the face recognition system with the face image printed on paper),screen display attack(i.e.cheating the face recognition system by displaying the face video,photo or three-dimensional model on the digital screen)and mask attack(i.e.the pretender puts on the mask to cheat the face recognition system).In practice,the above attack methods will also load a variety of skills,such as bending the printed face to make it have a general three-dimensional structure of the face;There are also printed faces with hollowed out eyes and mouth covered on the fake face(actually equivalent to a simple mask),so as to integrate the movement of key components of the real face into the fake face.Generally speaking,a good face counterfeiting technology aims to present realistic apparent texture,accurate 3D face structure,reasonable face motion and discriminative target identity features.Face recognition technology is widely used in various fields of life because of its practicability.In order to solve the security problem of face recognition system,in vivo detection has gradually become the focus of researchers.Aiming at the problem of insufficient discriminant information extraction caused by ignoring feature mining in most living detection algorithms,an adaptive face living detection network based on multimodal feature fusion is proposed in this paper.The work of this paper is summarized as follows:1.In this paper,a lightweight network combining gradient texture and group receptive field features is proposed.For the input with RGB image,the gradient information of the image is extracted through the gradient texture branch,the multi-scale spatial and semantic features are obtained by using the group receptive field branch,and then they are spliced and fused to make the network learn more abundant features.In addition,in order to improve the robustness of the liveness detection,the multiple supervision strategy of depth map and binary mask is used in this paper.Finally,the prediction results obtained by the depth map generator and the mask generator are added,and the prediction score is compared with the threshold to judge whether it is an attack face.The deceptive face attack simulated by the algorithm in Oulu NPU dataset is better than the existing attack algorithms based on feature extraction and auxiliary supervision.2.In order to meet the application requirements of face living body detection deployed to mobile device,improve the generalization of face liveness body detection algorithm and reduce the parameter scale,an improved adaptive face living body detection algorithm(Auto-FAS)based on neural architecture search algorithm(NAS)is designed in this paper.Considering that the input face information has different levels of features,multi-level Auto-FAS module is introduced to extract these features respectively.In order to guide the model to learn better face recognition ability and improve the training efficiency of the algorithm,this paper designs a pixel level binary supervision and improves the one shot training strategy.The algorithm is superior to the traditional neural architecture search algorithm according to the experiment results simulated in Oulu NPU dataset,and it is also significantly better than the face living detection algorithm based on auxiliary supervision in the cross dataset test.3.In order to verify the adaptability of multimodal feature fusion algorithm and adaptive neural architecture search algorithm,an adaptive face liveness detection system based on multimodal feature fusion is designed,and the feasibility of this algorithm is successfully verified.