Research On Detection Method Of Recording Replay Attack Based On Hybrid Feature And Improved Random Forest

Automatic speaker verification system has a good development and application in many fields because of its low equipment requirements,wide range of information collection and high user acceptance.However,at the same time of its rapid development,its security also faces huge risks.Recording and playback attack is a key problem.Different from other attacks,this attack only needs one recording device and does not need voice related knowledge,which greatly increases the possibility of recording and playback attack.Therefore,it is more and more important to improve the security of speaker verification system and realize the security protection based on recording and playback attack,which is also the research content of this paper.In order to detect playback voice better,the work and innovation of this paper are mainly divided into the following three parts.(1)In the detection of recording playback attack,appropriate features can better highlight the difference between real voice and playback voice.Therefore,this paper proposes a new feature LCSVD by further analyzing the spectrum.Besides,aiming at the problem of low stability of IMFCC features in different speech,M-IMFCC is proposed.In GMM model,these two features have better classification performance than traditional features when equal error rate is used as evaluation standard.Compared with the baseline system based on CQCC,the performance of LCSVD in the development set and the evaluation set is improved by 66.8% and 47.8% respectively,and the performance of M-IMFCC in the development set and the evaluation set is improved by 59.8% and 8.9% respectively.In addition,in order to reduce the distortion of traditional MVN caused by high-dimensional parameter regularization,WMVN is introduced to improve the robustness of the system.Compared with MVN,the performance of LCSVD and M-IMFCC under the evaluation set is improved by 26.26%and 4.93% respectively(2)In order to further enhance the relevant information contained in the features,different features are fused to form mixed features.LCSVD and M-IMFCC have good recognition performance for playback speech in low frequency and high frequency respectively,but single feature will also ignore the key information in other frequency bands.In order to better represent the difference between real and playback voice,this paper uses GMM supervector and relief algorithm to fuse LCSVD and M-IMFCC.We also use random forest as a classifier to compare the impact of different dimensions of GMM supervector on the results,and find that the mixed features have good classification performance on the development set in different cases.(3)In order to achieve better classification detection,this paper combines genetic algorithm and random forest to build a hybrid model,Weighted Random Forest with Genetic Algorithm.The model effectively reduces the size of random forest,improves the overall efficiency of the system and the stability of the results.The equal error rate of the detection system based on hybrid feature and Weighted Random Forest with Genetic Algorithm is 1.59% in the development set and 21.62% in the evaluation set.Compared with RF,Ada Boost and GDBT,the performance of development set is improved by 13.11%,65.28% and 53.10% respectively.The performance on the evaluation set was improved by 4.97%,10.55% and 9.54% respectively.When the detection system is applied to ASV,its security is improved obviously.The experiment shows that the equal error rate of ASV system is reduced from 11.72% to 7.91%.