Description:
Source mobile recording device identification aims to identify the source device through intrinsic audio characteristics, and it has been widely employed in various digital audio forensic scenarios. However, the existing methods have the problem of separating the short-term spatial information and long-term temporal information in the audio signal, resulting in insufficient utilization of source device feature/characteristic information and limited recognition accuracy. To solve this problem, we proposes a source mobile recording device identification method based on feature representation learning, which includes sequential gaussian mean matrix (SGMM) feature extraction and spatio-temporal deep learning model establishment, SGMM can effectively represent spatio-temporal information on a single feature and and the corresponding spatio-temporal model we have built can also be matched with it. For SGMM feature extraction phase, we first extracting the Mel-Frequency Cepstral Coefficients (MFCC) in the audio as acoustic features, and then dividing them into chronological feature segments to extract the SGMM by building a gaussian mixture model (GMM). Futhermore, We propose a deep spatio-temporal model to focus on the short-term and long-term information in the features respectively. We construct a convolutional neural network (CNN) to perform representation learning on SGMM features to extract the deep bottleneck features, and establish the corresponding bi-directional long short-term memory network (BiLSTM) to extract the temporal information in the sequential bottleneck features, achieved simultaneous analysis of the spatial complexity and temporal complexity of the source device signal. The Softmax classifier is used for the classification judgement, we calculate the cross entropy of each device category for identification. The experimental results show that the method can effectively identify a large number of recording device models, and adequately take into account the characteristics of digital audio itself, identification performance better than state-of-the-art results