End-to-End Mobile Phone Recognition Based On Deep Representation Learning

With the rapid development of digital media,many easy-to-use digital media editing software has been derived.However,the popularity of editing software makes it difficult to distinguish the authenticity of digital media.If lawbreakers use this software,it will seriously disrupt the social order.Therefore,it's essential to judge the authenticity,integrity,and origin of digital media.And the research content in this paper is audio-based mobile phone' recognition,a subcategory of digital media forensics.Mobile phone is currently the most important digital media devices.And audio has broad prospects in the news,justice,military,and security fields.So audio-based mobile phone recognition is an urgent research topic.The traditional mobile phone identification method extracts the audio time domain,fre-quency domain,or cepstrum domain features as the mobile phone's inherent features.Then the mobile phone is identified through an appropriate characterization model.Different traditional methods have various advantages and disadvantages,and it is difficult to com-pare them uniformly.Therefore,the universal background model-Gaussian mixture model(UBM-GMM),a widely used mature mobile phone identification method,is chosen as the baseline in this paper.In the UBM-GMM method,UBM is first built to represent mobile phones' common features in audio information.And other GMMs are constructed by opti-mizing UBM parameters based on adaptive algorithms.Every GMM represents one brand and one model of a mobile phone.And maximum likelihood estimation algorithm is used to predict the result of mobile phone recognition.But this method has three problems,manual intervention,feature extraction,and computational cost problems.And in deep learning methods part,many proposed works are compared with CNN method.Thus,CNN has been experimented many times.Its performance and generalization can be proved.So CNN is set as the other baseline in this thesis.The disadvantage of CNN method is there are many better methods to deal with long dependence problem in audio.In this experiment,CNN is extracted based on MFCC features.And CNN structure is a basic four convolution layers structures.In this thesis,three mobile phone methods are illustrated.The first method,based on the representation learning mobile phone recognition method,is proposed to overcome the dis-advantages of UBM-GMM.This method extracts the deep feature of segments of audio files by CNN.And then,deep features of segments from the same file are spliced.Finally,this spliced feature matrix is classified by the Bi-LSTM algorithm.But the performance of the first method is not prominent enough,and the segmentation in the first method is very ran-dom.So deep feature fusion method is proposed to improve the performance of recognition.This method fuses two different deep feature s together to enhance the accuracy of represen-tation features.However,the structure of this method is complicated.And ResNet,which extracts one component of features,doesn't highlight its advantages.Considering the ex-cellent performance and good plasticity of LSTM,a recognition method with attention and stack Bi-LSTM is put forward.Audio's deep features are extracted by stack Bi-LSTM struc-ture,and deep features of weight redistribution represent a mechanical fingerprint of mobile phones.Compared with baselines,all proposed approaches have an improvement in con-struction and performance.And the third method has the best performance in mobile phone recognition.