Research Of Recording Devices Source Recognition Based On Deep Representation Learning

As one of the important branches in the field of passive digital audio forensics,recording device source recognition technology is widely used in judicial security,media forensics,military information and other fields.At present,the recording device source recognition technology based on traditional methods were fully developed.But there are still many key challenges to be solved in the recording device source recognition technology based on deep learning.This thesis mainly solves the following key problems:(1)the feature representation in the field of recording device source recognition is underdeveloped and the model contains too many parameters designed manually;(2)the feature information is single,ignoring the temporal information in the feature of the recording device source;(3)the training time of deep network model is too long.(1)Aiming at the problem of insufficient feature representation in the field of recording device source recognition and the model contains too many hand-designed hyperparameters.This thesis proposes an end-to-end multi-feature fusion recording device source recognition method.This method first uses MFCC,GSV and I-vector features as input,then extracts the depth representation features of MFCC and GSV through a deep network,and finally uses the attention mechanism to achieve effective complementary fusion of the three features.In addition,the method also uses an end-to-end recognition architecture,which can effectively solve the problem of traditional models containing too many manually designed hyperparameters,and make the model test results more stable.The recognition accuracy of this method is 97.4%,which is 5.8% higher than the benchmark model.(2)Aiming at the single information in the recording device source feature,the temporal information problem in the recording device source feature is ignored.this thesis proposes an end-to-end spatial and timing information fusion recording device source recognition method.This method uses the Bi LSTM network to extract the deep temporal features in MFCC,and DNN to extract the deep spatial features in GSV.Then the attention mechanism is used to realize the fusion of deep temporal features and deep spatial features.Compared with the multi-feature fusion recording device source recognition method,this method reduces a feature input,and the recognition accuracy is increased by 0.1%.(3)Aiming at the problem of long training time for deep networks,this thesis proposes a recording device source recognition method based on the fusion of dense residual TCN temporal and CNN spatial information.This method uses spatial,temporal and channel attention mechanisms respectively.The spatial and temporal attention mechanisms assign weights to GSV and MFCC respectively,which enhances the characterization of the two features.The channel attention mechanism assigns weights to depth temporal and spatial features,which can effectively merge the two features.In addition,this method proposes a dense residual TCN to extract deep temporal features.Compared with the DNN deep space feature and Bi LSTM deep temporal feature fusion method,the recognition accuracy of this method is increased by0.1%,and the training time is reduced by 20 times,and improves the overall recognition efficiency of the recording device source recognition task.In summary,this thesis proposes an end-to-end multi-feature fusion device source recognition method based on the problem that the source identification feature representation of recording device is not strong enough and the model contains too many hand-designed hyperparameters.Secondly,in order to solve the problem of single source feature information of recording device,this thesis proposes an end-to-end spatial and temporal feature fusion recording device source recognition method.Finally,in order to solve the problem of too long training time for the source recognition task of recording device,this thesis proposes a source recognition method for recording device based on the fusion of TCN temporal features and CNN spatial features based on dense residuals.