数字音频被动取证关键技术研究

发布时间：2018-01-10 06:29

  本文关键词：数字音频被动取证关键技术研究　出处：《宁波大学》2016年博士论文　论文类型：学位论文

  更多相关文章： 数字音频 被动取证 来源识别 压缩历史检测 篡改定位 隐写分析

【摘要】：数字音频是人们日常生活中最容易获得数字媒体之一。除了以购买、下载的方式获得音频文件外,还可以通过实时录制的方式生成音频/语音文件。然而,音频编辑和处理软件的不断发展和完善,使得对音频的编辑和修改变得更加简单和廉价。同时,人耳也很难察觉这种修改留下的痕迹。因此,如何有效验证数字音频的原始性、完整性和真实性,就成为了数字音频被动取证技术迫切需要解决的问题。本文对数字音频被动取证中的关键问题和技术进行研究和探索,主要在取证音频数据库构建、音频来源取证、音频压缩历史检测、音频内容篡改检测及音频隐写分析这五个方面开展了研究工作:1.针对目前数字音频取证领域基准音频/语音库缺乏的问题,本文分别以CD音频抓轨和现场语音录制的方式,构建了一个基础音频数据库(CKC-AD)和一个基础语音数据库(CKC-SD)。前者包含2种类型,以及超过5种时长、10种音乐流派、4种语言的音频文件,共11172个;后者使用38种不同型号录音设备、对31个(21男10女)说话人分别录制了朗读和口语两部分语音。另外,本文在CKC-SD的基础上,依据具体研究内容,进一步构建了TIMIT翻录语音库、二次翻录音频库和设备本底噪声数据库。2.本文音频来源取证方面的工作由二次翻录音频检测和录音来源设备识别两部分组成:针对目前二次翻录音频检测方法仅涉及单一偷录或回放设备的问题,本文深入分析了音频回放翻录过程中不同偷录和回放设备对二次翻录音频的影响,并根据二次翻录音频和原始录制音频在高频信息量分布上的差异构建了特征向量。实验结果表明,该方法能有效区分原始录制音频和二次翻录音频,综合分类准确率达到了98.47%。另外,将该方法集成到GMM-UBM说话人识别系统中,可大幅提高其抵抗音频回放攻击的能力,使其等错误概率(EER)降低了47.06%。针对目前大多数录音来源设备识别方法均是基于美尔倒谱系数(MFCC)特征或其他声学特征的思路,本文从录音设备本身的特性切入,提出了两种录音来源识别的方法。方法一是利用不同型号设备在音频编码过程中对各编码参数使用特点的不同,构建相关的统计量特征实现录音来源设备的识别。实验结果表明,该方法对CKC-SD中10款录制MP3音频设备的平均识别率为99.97%,对14款录制AAC/M4A音频设备的平均正确检测率为96.53%。另一个方法对方法一受录音格式限制的局限性进行了改进。在深入研究不同录音设备本底噪声的基础上,提出了设备本底噪声的估计方法,并针对估计的本底噪声构建了频谱形状特征和频谱分布特征来表征各设备。该方法实现了对CKC-SD库中34款设备较为准确的区分,其平均分类准确率为95.53%。3.针对目前涉及较少的AAC音频双压缩检测,本文提出了一种基于Huffman码表索引的双压缩检测方法。通过分析双压缩操作对码表索引分布的改变,统计了码表索引的直方图和Markov单步转移概率作为分类特征。对低转高码率的双压缩音频(FAAC/FAAD2编解码器),检测准确率达到了99%以上;但在相同码率情况下,分类准确率仅为79.56%。与该领域典型方法的对比结果表明,本方法整体上检测准确更高。另外,对MP3音频的压缩历史检测(不超过3次)和码率估计进行了探索,本文研究了Huffman码表索引和比例因子在多次压缩情况下的渐进式变化,有针对性地构建了均差、概率分布和互相关性统计量组成特征向量。实验结果表明:本方法对双压缩MP3音频的检测准确率较目前该领域的几种典型方法,整体上有所提升;在三次压缩检测方面,对低转高、相同码率及高转低码率的情况(前提条件:BR2=BR3),分类准确率分别为97.73%、94.56%和80.28%,另外,在第三次码率高于128kbps时,能较为有效地从一、二、三次压缩音频混合集中区分三者。4.针对常见的篡改操作,本文提出了两种篡改定位的方法。方法一受帧偏移方法的启发,利用篡改前后音频量化特性的不一致性,将量化前后小值频率系数的转化率作为检测变量实现篡改定位。实验结果表明,该方法对192kbps(原始未篡改MP3音频的码率)及以下音频的篡改定位准确率达到了98%。但该方法仅对篡改后以非压缩格式保存的音频有效。方法二基于重压缩对帧结构被破坏部分的音频具有校正功能的原理,发现了篡改前后的音频片段在估计的压缩次数上的不一致性,从而将这种不一致性用于篡改定位。虽然从实验结果来看,由于受限于双压缩检测方法的精度,该方法的定位准确率暂无法令人满意,但为研究压缩音频的篡改检测开辟了一种新的思路。另外,该方法实用性更强,可检测篡改后的双压缩音频。5.针对MP3Stego低嵌入率情况下检测准确率不高的问题,通过分析MP3Stego隐写操作对MP3音频量化频谱系数的影响,有针对性地对量化频谱系数幅值的差值构建了块内和块间的Markov单步转移概率特征,实现了对低嵌入率下MP3Stego的有效检测。实验结果表明,该方法对嵌入强度为10.6%的MP3音频,平均检测准确率能达到90.74%。随着码率的降低,检测性能会有所下降,但仍优于现有的典型方法。另一方面,本文还对另一个MP3隐写工具——Under MP3Cover的隐写原理进行了深入剖析,发现其嵌入方法的核心是连续的LSB替换,但嵌入的位置间隔是通过参数Bit Spacing控制。依据其隐写原理,对RS分析法进行了改进,成功实现了对Under MP3Cover的检测,并能有效估计嵌入秘密信息的长度。另外,对改进方法中最佳翻转算子的选择、是否重叠分组以及参数Bit Spacing对嵌入强度估计准确性的影响等问题进行了讨论与分析。
[Abstract]:Digital audio digital media is one of the most easily available in people's daily life. In addition to the purchase, Download way audio files, but also through the real-time recording mode to generate audio / audio files. However, the development of audio editing and processing software and improve the audio editing and revising easier and cheap. At the same time, the human ear is difficult to detect the modified traces. Therefore, how to effectively verify the original digital audio, integrity and authenticity, has become a digital audio forensics problem need to be solved urgently. This paper studies and explores the key issues of digital audio forensics in and technology. The main building in the audio database forensics, audio forensics, audio compression history detection, audio content tamper detection and audio steganalysis which carried out the five aspects of research work: 1 At present, digital audio forensics field reference audio / speech database in this paper, the problem of lack of CD audio ripping and field voice recording mode, constructs a basic audio database (CKC-AD) and a basic speech database (CKC-SD). The former includes 2 types, as well as more than 5 long, 10 music genre, 4 language audio files, a total of 11172; the latter uses 38 types of recording equipment, of 31 (21 male and 10 female) were recorded speaker reading and spoken two part of speech. In addition, based on CKC-SD, according to the specific contents, further build TIMIT rip speech database two times, recording frequency database and background noise of the.2. database of the audio source forensics work consists of two parts: two rip audio detection and recording source device identification: according to the current two times audio frequency detection method only involved And single Toulu or playback equipment, this paper analyses the influence of the two recording and playback equipment of different audio playback audio ripping to steal the ripping process, and construct the feature vector according to the difference of the two rip audio and original recording audio in high frequency information of the distribution. The experimental results show that this method is effective the distinction between the original recording audio and two audio dubbing, comprehensive classification accuracy rate reached 98.47%. in addition, the scheme is integrated into the GMM-UBM speaker recognition system, can greatly improve the ability to resist the audio replay attack, so as to reduce error probability (EER) of 47.06%. for most of the recording source device identification methods are Mel based on cepstral coefficients (MFCC) or other characteristics of the acoustic characteristics of ideas, this essay starts with the characteristics of the recording equipment itself, proposed two methods of recording source recognition method is. By using the characteristics of the encoding parameters in the audio encoding process in different types of different equipment, construction of statistic features related to the recognition of the recording source device. The experimental results show that the method of CKC-SD in 10 recorded MP3 audio equipment average recognition rate is 99.97%. The average correct detection of AAC/M4A audio recording of paragraph 14 of the improvement rate of 96.53%. another method for a limitation of the recording format restrictions. Based on a thorough study of the different recording equipment of background noise, proposes a new estimation method of background noise of equipment, and for the estimation of the background noise of the constructed spectral shape feature and spectrum distribution to characterize the equipment. This method can distinguish the CKC-SD library on the 34 devices more accurately, the average classification accuracy of 95.53%.3. currently involves less AAC audio double compression detection in this paper. We propose a double compression detection method of Huffman code based on index. Through the analysis of double compression operation on the table index distribution, the statistical table index histogram and Markov single step transition probability as classification feature. On the double compressed audio low to high rate (FAAC/FAAD2 codec), the detection rate has reached more than 99%; but at the same rate, the classification accuracy rate is 79.56%. and the field only compare the typical method. The results show that the method of the overall detection accuracy is higher. In addition, the detection of MP3 audio compression history (not more than 3 times) and rate estimation are explored, this paper studies the Huffman index and the scale factor in the table multiple compression incremental change situation, targeted to build the mean probability distribution and correlation statistics form the feature vector. The experimental results show that this method of dual MP3 audio compression The detection accuracy is currently in the field of several typical methods, improved on the whole; in the three compression test, turn high to low, the same rate and high to low bit rate (prerequisite: BR2=BR3), the classification accuracy rate were 97.73%, 94.56% and 80.28%, in addition, in the third rate more than 128kbps, can be more efficiently from one, two, three times the compressed audio mixed centralized distinction between the three.4. for the common tampering, two methods are proposed in this paper. A method of tamper localization inspired by frame offset method, inconsistency in the use of tamper audio characteristics before and after quantization, transformation before and after quantization the small value of frequency coefficient detection rate as the variable to achieve tamper localization. The experimental results show that the method of 192kbps (the original rate untampered MP3 audio and audio) tamper localization accuracy reached 98%., but this method only for tampering with non pressure Save the audio formats effectively. Methods two based on compression principle has correction function of frame structure was destroyed part of the audio, audio clips found before and after tampering in the estimation of the number of compression inconsistency, and this inconsistency for tamper localization. Although from the experimental results, due to the limited double compression detection accuracy, positioning accuracy of this method will not be satisfactory, but opens up a new way to study the compressed audio tampering detection. In addition, this method is more practical and can detect tampering after double compressed audio.5. for MP3Stego detection under the condition of low embedding rate is not high accuracy rate, through the analysis of MP3Stego steganography effects on MP3 audio quantization of the spectral coefficients, for quantitative spectral coefficients of amplitude difference of building block and block the Markov single step transition probability. Sign, realize the effective detection of MP3Stego under low embedding rate. The experimental results show that the method of embedding strength for 10.6% MP3 audio, the average detection rate can reach 90.74%. as the rate decreased, the detection performance will decline, but still better than the existing typical methods. On the other hand, in-depth analysis in this paper also write tools -- Under MP3Cover on another MP3 hidden hidden principle, found the core embedding method is continuous LSB replacement, but the position is controlled by embedded interval parameter Bit Spacing. On the basis of the principle of steganography, RS analysis method was improved, the successful implementation of the detection of Under MP3Cover. And can effectively estimate the embedding length. In addition, the best method in selection of flip operator, whether overlapping grouping and parameters of Bit Spacing on the embedding strength influence the accuracy of the estimation results and other issues are also discussed. Theory and analysis.

【学位授予单位】：宁波大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP309
，

本文编号：1404249