【作者】 王静；

【导师】 王志锋；

【作者基本信息】 华中师范大学 ， 数字媒体技术， 2019， 硕士

【摘要】 数字音频数据易于保存、编辑和传播的优点给日常生活带来诸多便利和乐趣,但是同时让一些不法分子有机可乘,他们可以对数字音频进行恶意地篡改并进行广泛传播,且仅凭感官难以察觉出来。若将这样的音频文件应用在法庭录音举证、虚假新闻传播等场合下,可能会损害法律公正和社会信任秩序。因此,验证数字音频的真实性和完整性的数字音频篡改被动检测研究显得尤为重要。本文是对数字音频篡改被动检测关键技术的研究,通过分析数字音频篡改被动检测技术要解决的问题和任务,进行三个方面内容的研究:1)基于先验信息的数字音频篡改检测,2)无先验信息的数字音频篡改检测,3)数字音频的篡改点定位。从下面几个层面切入,在特征层面,结合音频信号自身的属性,分析不同的篡改特征的检测效果,通过特征融合技术获取最优的特征表达。在模型层面,尝试借鉴相关领域的研究成果,建立语音信号特征模型,构建语音信号与模型之间的映射关系,实现无先验信息的盲篡改检测。在实现系统层面,分析传统算法不足,降低算法复杂度,提高自动化程度,实现自动化与多层次的数字音频被动篡改检测分析系统。本文主要工作如下:(1)针对已知可能的篡改类型,判断待测信号是否存在这种篡改的情况,进行基于先验信息的数字音频篡改检测研究。首先本文提出基于电网频率波动一致性的数字音频篡改自动检测方法,基于ENF相位和频率波动的稳定性,应用DFT0方法和Hilbert变换分别估计ENF成分的相位和瞬时频率,并构建统计特征,应用优化的SVM算法对待测信号真伪进行分类决策,分类准确度达到92.2%。其次,在前一算法的基础上,为了充分挖掘ENF的波动特征,提出基于特征融合的深度学习自动分类方法,使用谱线拟合特征能够更好表征ENF的局部突变情况,基于并行策略特征融合方法DCA减少特征维数后,应用深度随机森林进行分类决策,分类准确度达到95.2%。相比传统算法,所提算法的鲁棒性、泛化性能均得到提高,且实现了篡改判定的自动化。(2)针对盲篡改检测,即没有任何先验信息的仅根据音频本身的特点来判断待测信号是否遭到某种形式的篡改或处理的情况,进行无先验信息的数字音频篡改检测研究。首先,本文提出基于ENF-UBM的数字音频篡改检测方法,建立通用背景模型模拟原始语音特征的分布,再用少量的目标数据库信号进行自适应,设定评分函数量化待测信号与目标模型的匹配度,基于ROC曲线寻找最佳阈值,以此区分待测信号的真伪,其等错误率降到10%左右。其次,为了优化决策依据,消除其中经验性阈值设定,提出基于电网频率波动超矢量的数字音频盲篡改检测方法,为每个待测信号自适应目标模型,通过理论分析将模型映射成特征超矢量,以特征超矢量的分类结果做为决策依据,实现方案的自动化,分类精度达到86%。该方法取得了检测精度与检测过程自动化的平衡,但相较于传统的启发式算法,其更具有更好的鲁棒性。(3)针对篡改信号的篡改区域定位困难的问题,进行数字音频篡改点定位研究。本文将篡改点的位置分为两种情况。篡改点位于静音段时,本文将传统BIC方法改进为采用迭代的方式对篡改点存在与否进行模式化分析,首先是粗判断得到可疑点截断窗,对截断窗的BIC值序列进行曲线拟合,将拟合参数进行分类,判断为存在篡改点的阶段窗再进一步确定篡改点的位置。本方案在减小计算量的同时实现了篡改点的精确定位,分类精度达到99.7%。针对篡改点相距较近的情况,应用更加复杂的特征选择模型即GMM-BIC进行特征建模,使之能够更加准确地模拟短时语音的特征分布,同样的操作方式下分类精度达到90%。篡改点位于语音段时,借鉴语音识别和数字图像取证等相关研究领域的成果,结合数字音频信号自身的特性,将音节片段转换为音节语谱图,量化音节之间的相似度,制定决策依据,判断篡改信号中是否存在两个相似的音节,相比其他音节描述特征方法,本文方法具有相对较少的计算量和较好的鲁棒性。(4)根据数字音频篡改被动检测任务需求和检测流程,进行数字音频被动取证系统的研发,证明本文所提算法的有效性,实现自动化与多层次的数字音频篡改被动检测分析系统。本研究引入图像取证、机器学习、特征融合等相关研究理论,改善了传统检测方法成效,完善了数字音频篡改被动检测的理论研究,可以应用在司法取证、知识发现、版权保护等众多需要验证数字音频真实性和完整性的场合。更多还原

【Abstract】 The advantages of easy preservation,editing and dissemination of digital audio data bring many conveniences and pleasures to our daily life,but at the same time,some illegal elements can take advantage of it.They can maliciously tamper with digital audio and spread it widely,and it is difficult to detect it only by their senses.If such audio files are used in court recording and proof,false news dissemination and other occasions,it may damage the legal justice and social trust order.Therefore,it is particularly important to verify the authenticity and integrity of digital audio tampering passive detection.This paper studies the key technologies of passive detection of digital audio tampering.By analyzing the problems and tasks to be solved in passive detection of digital audio tampering,three aspects are studied:1)digital audio tampering detection based on prior information,2)digital audio tampering detection without prior information,3)tamper location of digital audio.From the following aspects,at the feature level,combined with the attributes of audio signal itself,the detection effect of different tampering features is analyzed,and the optimal feature expression is obtained by feature fusion technology.At the model level,we try to draw lessons from the research results in related fields,establish the speech signal feature model,construct the mapping relationship between the speech signal and the model,and realize blind tamper detection without prior information.At the system level,the traditional algorithms are insufficient,the complexity of the algorithms is reduced,the degree of automation is improved,and the automatic and multi-level digital audio passive tamper detection and analysis system is realized.The main work of this paper is as follows:(1)In view of the known possible tampering types,the digital audio tampering detection based on prior information is studied to determine whether there is such tampering in the signal to be tested.Firstly,this paper presents an automatic detection method of digital audio tampering based on the consistency of power grid frequency fluctuation.Based on the stability of ENF phase and frequency fluctuation,DFT0 method and Hilbert transform are used to estimate the phase and instantaneous frequency of ENF components respectively.Statistical characteristics are constructed.The optimized SVM algorithm is applied to classify the true and false signals,and the classification accuracy reaches 92.2%.Secondly,on the basis of the previous algorithm,in order to fully mine the fluctuation characteristics of ENF,a deep learning automatic classification method based on feature fusion is proposed.The spectral line fitting feature can characterize the local mutation of ENF better.After reducing the feature dimension based on parallel strategy feature fusion method DCA,the classification accuracy of deep random forest is 95.2%.Compared with the traditional algorithm,the robustness and generalization performance of the proposed algorithm are improved,and realize the automation of tamper detection.(2)Aiming at blind tamper detection,that is,to judge whether the signal under test has been tampered or processed according to the characteristics of audio itself without any prior information,digital audio tamper detection without prior information is studied.Firstly,this paper proposes a digital audio tampering detection method based on ENFUBM,establishes a general background model to simulate the distribution of the original speech features,adapts a small number of target database signals,sets a score function to quantify the matching degree between the signal to be measured and the target model,and finds the optimal threshold based on ROC curve,so as to distinguish the true from the false of the signal to be measured,and its equal error rate is reduced to about 10%.Secondly,in order to optimize the decision-making basis and eliminate the empirical threshold setting,a digital audio blind tampering detection method based on the frequency fluctuation supervector of power grid is proposed.For each signal to be measured,the model is mapped to the feature supervector through theoretical analysis,and the classification result of the feature supervector is used as the decision-making basis to realize the automation of the scheme.The classification accuracy reaches 86%.The method achieves a balance between the detection accuracy and the automation of the detection process,but it has better robustness than the traditional heuristic algorithm.(3)Aiming at the problem that the tampering area of the tampering signal is difficult to locate,the digital audio tampering point positioning research is carried out.In this paper,the location of tamper points is divided into two cases.When the tamper point is located in the silent segment,this paper improves the traditional BIC method to model the existence of tamper point by iteration.Firstly,the truncated window of the suspicious point is obtained by rough judgment,the BIC value sequence of the truncated window is fitted by curve fitting,the fitting parameters are classified,and the stage window with tamper point is judged to determine the location of the tamper point further.This scheme reduces the computational complexity and achieves the accurate location of tampered points,with the classification accuracy of 99.7%.In view of the fact that the tamper points are close to each other,a more complex feature selection model,GMM-BIC,is used to model the features,which can more accurately simulate the feature distribution of short-term speech.Under the same operation mode,the classification accuracy reaches 90%.When the tamper point is located in the speech segment,referring to the achievements of the related research fields such as speech recognition and digital image forensics,combining with the characteristics of digital audio signal itself,the syllable segment is converted into syllable spectrogram,the similarity between syllables is quantified,and the decision-making basis is made to determine whether there are two similar syllables in the tampered signal.Compared with other syllable description methods,this method has the advantages of less computation and better robustness relatively.(4)According to the requirement of digital audio tampering passive detection task and detection process,the research and development of digital audio tampering passive forensics system proves the effectiveness of the proposed algorithm,and realizes automatic and multi-level digital audio tampering passive detection and analysis system.In this study,image forensics,machine learning,feature fusion and other related research theories are introduced to improve the effectiveness of traditional detection methods,improve the theoretical research of passive detection of digital audio tampering,which can be applied in judicial forensics,knowledge discovery,copyright protection and many other occasions that need to verify the authenticity and integrity of digital audio.更多还原