Research On Intelligent Terminal Security Based On Face Anti-spoofing

Face recognition technology is applied to the mobile phone of tablets and other mobile intelligent terminal is becoming more popular.However,printing pictures,replaying video and making masks can forge face recognition mechanism of mobile intelligent terminal,which poses a great threat to its identification system.Therefore,how to detect the false face is a hotspot of mobile intelligent terminal face anticounterfeiting technology.Traditional face anti-counterfeiting technology in hand-crafted features as the basis of authenticity that makes one face different from another,because of the added artificial restrictions,usually applied to prevent printing photos attacks model is difficult to to resist replay attack,the video there are for different kinds of poor universality problems in order to improve the algorithm of generality,deep learning is used in face security domain.However,there are three major problems with existing models based on deep learning: First,poor adaptability to changes in facial posture and lighting conditions.Secondly,there is a lack of clear supervision information.What the model learns is not the key feature to distinguish the human faces,which leads to the low accuracy of detection.Third,it is sensitive to face video collection equipment and methods,and poor generalization of different data sets.To solve the above problems,this paper proposes a face authenticity detection scheme combining static features and dynamic features.The main work and innovation points are as follows:1.The face authenticity detection scheme combining static features and dynamic features proposed in this paper firstly collects face video through the camera of mobile intelligent terminal,and then samples the continuous frames.The depth map of the first frame is extracted as the static feature.Then the optical flow guidance feature is introduced and the dynamic features are obtained by analyzing the dynamic changes of human faces in all frames after sampling.The solution is to use the fusion coefficient lambda to control the relative importance of static features or dynamic features.It integrates the advantages of static features and dynamic features,making the fusion features have strong universality and generalization at the same time.2.In view of the existing algorithms adapted to different face postures,expressions and light conditions are poor and the problem of detecting accuracy is not high,this paper puts forward a kind of based on 3 d point cloud image and depth map of static face feature extraction method,the reconstruction of 3 d point cloud picture first,then based on the extracted face depth map in order to enhance to different facial expressions and the adaptability of light conditions,the method adopts the PRNet algorithm,by setting up special UV space mapping,in the face of 3 d point clouds reconstruction images In order to extract features with high degree of discrimination for true and false faces,and thus improve the accuracy of detection,this method normalizes 3D point cloud maps,which are used as real markers to guide the deep network model for training,introduces explicit supervision information,and extracts the depth maps of faces as static features.3.In view of the existing algorithms for different data sets with poor generalization problem,this paper proposes a mechanism based on temporal relationship and attention of the face is the method of dynamic feature extraction method using optical flow guide features consecutive frames in the face of short-term dynamic changes,said of the short-term dynamic changes by CGRU module accumulation get dynamic change for a long time,has given in successive frames of temporal relations at the same time,and keep the face of the space information;At the same time,the attention mechanism is used to weighted video frames at different time points according to the importance to obtain the dynamic features of the face,which has a strong generalization performance for different data sets.4.In order to verify the effectiveness of the scheme,this paper based on CASIA-MFSD,Replay Attack and SiW three data sets within the data set tests and cross test data sets,and the characteristics of single feature and fusion compares the experimental results show that the effect of face false detection based on static or dynamic characteristics of error rate is lower than the existing algorithm of error rate in different degrees.The static features are better in the data set internal test and the dynamic features are better in the data set cross test.Compared with the single static feature or dynamic feature,the fusion feature has a lower error rate,with the lowest average classification error rate(0.18±0.23)% in the internal test of the data set and the lowest semi-error rate 17.0% in the cross-test of the data set,which can be better used to distinguish face authenticity.