Facial recognition technology is a biometric identification technology used to identify and verify facial identities.It is widely used in various fields,such as security monitoring,mobile payment,and access control authentication.However,the security issues of facial recognition technology have also received widespread attention.Attackers can deceive facial recognition systems,thus accessing unauthorized information or resources,such as video replay attacks.To ensure the security of facial recognition systems,this dissertation studies video matching technology and proposes a method for detecting video replay attacks.The principle of this dissertation to solve the replay attack problem is to prevent authenticated videos from being resubmitted.Historical authenticated videos are stored and recorded,and new authenticated videos are matched,rejecting any video that can be matched.Implementing this method requires solving two problems:video feature extraction and feature index matching.The main research content of this article is as follows:(1)A feature extraction framework based on 3D convolutional neural networks is proposed to address the problem that current video similarity feature extraction methods cannot fully utilize the temporal dimension of videos.This method extracts the spatial and temporal features of the video end-to-end through 3D convolutional networks,representing the video with multiple segment features,improving the utilization of video information.In addition,this method also uses triplet loss to minimize the distance between attack samples and original videos as much as possible,enabling the model to learn more useful features and improve the accuracy of the model.(2)A method for constructing feature index based on online product quantization is proposed to address the problem that commonly used index construction methods for face recognition authentication videos cannot be dynamically updated in real-time.This method updates the quantization center based on the traditional product quantization method,maintaining the old index unchanged,and can effectively achieve dynamic updates of feature indexes.Combined with the longest path algorithm of directed acyclic graphs to achieve video segment matching,video replay attacks can be detected,improving the detection efficiency.(3)A video replay attack detection system is designed and implemented.The system is divided into a web client and a server and implements user management,video management,replay attack detection,and detection record management functions.The replay attack detection function uses the 3D convolutional feature extraction network and online product quantization index with directed graph matching algorithm proposed in this article for video matching.Test results show that the system can effectively detect video replay attacks. |