| Automatic Speaker Verification(ASV)is designed to verify whether the speaker's claimed identity is true.It is one of the widely used biometric technologies.Compared with biometric technologies such as fingerprints and faces,it has noncontact,no need for light and other advantages.However,with the increasing dependence and trust of the application system on speaker verification technology,how to prevent various attacks that may be faced is an important and difficult problem that the ASV application system must solve.Replay attack refers to impersonating the target speaker by replaying the stolen/recorded target speaker's voice,which is a common method of attacking the speaker verification system.The traditional ASV technology is difficult to distinguish this attack.Therefore,it is necessary to study special technical means to detect such attacks.Aiming at the problem of speaker replay attack detection,this paper focuses on researching effective feature extraction methods to achieve the goal of effectively detecting replay attacks.The main content of the paper is as follows:1)An embedded feature extraction method based on Time Delay Neural Network(TDNN)is proposed.The TDNN network is used to extract embedded features,the focus loss function is used to solve the data imbalance problem in the training data,and the final attack detection classification is performed with softmax.Compared with the traditional Gaussian Mixture Model(GMM)method based on the Inverted Mel Frequency Cepstral Coefficient(IMFCC)feature,this method has a minimum tandem detection function(Minimum Tandem Detection Cost)in the evaluation set.Function,min-t DCF)and equal error rate(Equal Error Rate,EER)two indicators were increased by 10.28% and 15.87%,and achieved good results.2)A method of combining multiple segment-level embedded features to improve the performance of replay attack detection is proposed.First,extract the embedded features that reflect the environment and device information,and fuse their TDNN embedded features,and then,improve the detection performance by fusing multiple segment-level embedded features.Experimental results show that the min-t DCF and EER indicators of the fusion feature have a relative improvement of 13.95% and 11.38%,respectively.3)A speaker-independent feature extraction method based on adversarial training is proposed.In order to reduce the speaker information contained in the extracted features as much as possible,thereby reducing the interference of the speaker information on the detection of replay attacks,an adversarial training mechanism is adopted so that the extracted embedded features do not contain speaker information as much as possible.Experimental results show that the detection performance of this method exceeds that of the previous two methods,and it can more effectively detect speaker replay attacks. |
