Research On Replay Attack Detection Of Automatic Speaker Verification By High Frequency And Bottleneck Feature

Nowadays,the application of biological authentication is more and more widely used,but the related research indicates that the biological verification technology is more vulnerable to malicious spoof attack.Although some research progress has been made in the research of electronic spoof detection,there are still many difficulties in this problem,and the biological authentication system is still easily affected by the spoof attack.As a typical biological verification system,the speaker verification system needs the technology of spoof detection to ensure its reliability and security.In recent years,the replay attacks spoof detection has become a hot topic in the field of spoof detection,and a new research method is urgently needed to solve the problem of replay attack spoof detection.In order to improve the performance of replay attack spoof detection,we will explore the detection methods based on acoustic signal processing and deep learning detection methods.The former method is to research the replay attack spoof detection in the speaker verification based on the high frequency acoustic feature extraction and the feature processing method on the ASVspoof2017 dataset,then improves the performance of the speaker verification system for replay attack detection.Meanwhile we will explain the characteristic of replay attack from the angle of the signal propagation.Based on the deep learning method,the research of replay attack spoof detection is carried out on the ASVspoof2017 dataset by extracting the bottleneck feature of constructed network and using ensemble learning method.In this paper,the characteristics of replay attack are studied deeply,and we propose a new feature named Correction Inverted Mel Frequency Cepstral Coefficient(CIMFCC)for replay attack detection by improving Inverted Mel Frequency Cepstral Coefficient(IMFCC).The improvement includes the application of blackman window function and mean variance normalization.The experiment shows that CIMFCC is an acoustic feature that can detect the spoof speech effectively,and the Equal Error Rate(EER)of detection system based CIMFCC has a relative reduction of 51.06% compared to the baseline system.In order to improve the high frequency discriminative information loss in acoustic signal processing,we propose a bottleneck feature based on deep learning.And the Convolutional Deep Neural Network is used as a feature extraction method to capture the high frequency discriminative information of the spectrum,and the bottleneck feature is generated by the bottleneck layer in CDNN.The bottleneck layer transforms the related information into a low dimension representation.So the bottleneck feature can be regarded as the low dimension nonlinear representation of the input feature.Meanwhile,the experiment shows that the bottleneck feature is a more effective detection feature for the spoof detection problem.And the model based on the bottleneck features shows the best performance which EER is 8.40% on the evaluation set.Besides experiments show that the ensemble learning model is more suitable for solving replay attack detection problems than simple machine learning models.