| Digital audio is a ubiquitous multimedia,which can be downloaded from the internet and captured by audio recorders,etc.However,the widespread availability of low-cost and sophisticated digital audio editing software poses a serious threat to audio security.Therefore,how to examine the originality,authenticity and integrity of digital audio has become an urgent problem for passive forensics of digital audio.This doctoral thesis focuses on the key issues and techniques of digital audio forensics.In particular,five relevant aspects has been conducted as follows.1.Due to the lack of lack of benchmark audio/speech database for the relevant research area,an audio database(CKC Audio Database,CKC-AD)and a speech database(CKC Speech Database,CKC-SD)have been constructed.The audio database consists of 11172 audio files including 2 types,5 durations,10 genres and 4 languages.The speech database is recorded by38 sound recorders and 31 speakers(21 males/10 females)who are asked to finish reading and spontaneous corpus,respectively.In addition,TIMIT Recaptured Database,Speech Recaptured Database and Device Self-noise Database are established based on the CKC-SD for evaluating the algorithms proposed in this thesis.2.For detecting the recaptured audio,an algorithm based on the weighted information of the high frequency components is proposed.During recording procedure,the impact for the recaptured audio made by the replay and eavesdropping devices is studied.Experimental results show that the proposed method can perfectly distinguish the recaptured audio from the original one,and the classification accuracy exceeds 98%.Moreover,this method shows a huge improvement for the GMM-UBM speaker verification system to resist replay attack,and the EER(Equal Error Rate)is reduced by 47.06%.Differ from the mainstream MFCC-based approaches,two novel methods for source device identification based on the intrinsic properties of recording devices are proposed.One method extracts the statistics from the characteristics of codec implemented in the devices.Experimental results demonstrate that the identification accuracies of this method can reach99.97% and 96.53% among 10 MP3-recorded devices and 14 AAC/M4A-recorded devices,respectively.To adress the limitation that the above-mentioned method cannot work on other format audios,another approach considering device self-noise as the fingerprint of devices is proposed.Self-noise is estimated from the near-silent segments of recording.Spectral shape features and spectral distribution features of device self-noise are extracted for classification.Experimental results show that the self-noise has the ability to identify 34 recorders of different models.3.By analyzing the distribution alteration of Huffman codebook index between single and double compressed audios,a Huffman codebook-based scheme is presented for detecting double compressed AAC audios which is merely involved in double audio compression.The histogram and Markov transition probability of Huffman codebook indices are utilized as the feature vector.For up-transcoded audios(FAAC/FAAD2),the detection accuracy of the proposed method achieves 99%,while for the audio transcoded with same bitrate,the accuracy drops to 79.6%.Additionally,this paper explores MP3 multiple(up to three times)compression detection.The mean difference,probability distribution and correlation are extracted according to Huffman codebook and scalefactor to compose the feature vector.Experimental results illustrate that the proposed method performs better compared with the state-of-the-art methods.For triple compression,the total classification accuracies of up-transcoded,normal-transcoded and down-transcoded are 97.33%,94.56% and 80.28%,respectively.Furthermore,this method is able to classify the single,double and triple audios in a mixture audio set.4.For audio tempering detection,two forgery locating algorithms are proposed.Inspired by Yang's method,the quantization inconsistence in the original and tempered segments of the audio is used for locating the forgery position.Experimental results show that the proposed method can accurately detect the tempered location.However,this method has a limitation that it would be invalid to the tempered audio resaved in a compressed format.The other scheme is developed based on the compression calibration idea.The audio segment after the tempered position is probably misclassified as the single compressed segment while the part before the tempered position could be recognized as the double compressed segment correctly.Though location accuracy of this algorithm is not satisfactory so far,a promising way for detecting forgery in compressed audio is present.5.MP3 Stego hiding modifies the correlation between the quantized frequency coefficients.In order to improve accuracy for discovering the MP3Stego-embedded audio at low embedding-rate,the Markov features of the coefficients are extracted to capture the subtlealteration.Experimental results indicate that this scheme outperforms the prior arts on detecting MP3 Stego with low payload.On the other hand,a steganalytic method is proposed for another typical audio steganography tool Under MP3 Cover.This method,which is modified according to RS algorithm,is successful in detecting Under MP3 Cover.Furthermore,it can estimate length of the secret message embedded in MP3 audios. |
PDF Full Text Request