基于ENF信号的数字音频篡改盲检测研究

发布时间：2018-10-31 12:31

【摘要】：数字多媒体设备的广泛应用使得数字音频的录制越来越方便，而各种音频编辑软件的出现也使得音频篡改越来越容易。为了验证数字音频的原始性、完整性和真实性，寻求可靠的音频篡改检测方法变得更加迫切。基于电网频率(ElectricNetwork Frequency, ENF)的数字音频篡改检测是近年来获得较多关注的一种取证方法，但现有的基于ENF的方法存在一定的不足，比如在判断篡改时需要ENF参考数据库，篡改定位精度不高等。针对这些不足，本文研究无需ENF参考数据库的音频篡改盲检测方法，主要工作和创新如下：（1）提出一种利用ENF信号的最大相关偏移(Maximum Offset for CrossCorrelation, MOCC)的音频篡改检测方法。将ENF信号划分子块并计算起始子块与各子块取得最大互相关值时的偏移量，即最大相关偏移，根据各子块MOCC的一致性判断音频篡改，并实现篡改区域定位和篡改类型估计。针对不同的应用场景，还分别提出视觉检测法、自动检测法以及快速检测法。实验表明，该方法能较准确地定位篡改，并根据定位的篡改区域内容是静音部分还是语音部分，判断篡改类型是删除还是插入。另外，由于该方法是在时域进行检测，不需要转换到变换域，因此其计算量比现有的基于ENF相位的方法要小。（2）针对（1）中的方法在计算MOCC时受到噪声干扰的问题，提出一种改进的MOCC的音频篡改检测方法。引入一个理想的正弦波作为参考信号，利用参考信号与ENF各子块信号的多重互相关来增强ENF信号，并利用参考信号计算增强后的ENF各子块信号的MOCC，根据MOCC的变化情况来定位篡改及判断篡改类型。实验表明该方法能有效提高ENF信号的信噪比，减小MOCC受到的干扰，并能改善低虚警率情况下的篡改检测准确率。另外，在抵抗加性噪声、音频重采样和音频压缩方面也有一定的鲁棒性。（3）提出一种基于最小平均幅度差偏移(Minimum Offset for AverageMagnitude Difference, MOAMD)的双机制音频篡改检测方法，以提高篡改定位的精度。利用MOAMD来检测音频篡改，以简化MOCC的计算；采用双机制判断准则，即联合MOAMD的曲线变化情况和MOAMD曲线极值点的斜率变化情况，共同定位篡改区域。实验表明，该方法能有效提高篡改定位的精度，而且双机制中的机制二也能推广应用到其它方法。（4）提出一种基于ENF邻域相关系数的快速横向滤波(Fast Transversal Filter,FTF)的音频篡改检测方法，以提高现有方法的检测准确率。计算ENF邻域子块的相关系数，并对相关系数进行FTF自适应滤波，，根据滤波后的误差能量的变化情况来判断篡改，并实现篡改定位。实验表明，该方法能够有效提高篡改检测的准确率，在ENF波动范围较大和信噪比较低的情况下其优势更加明显。
[Abstract]:The wide application of digital multimedia equipment makes the recording of digital audio more and more convenient, and the emergence of various audio editing software makes audio tampering easier and easier. In order to verify the originality, integrity and authenticity of digital audio, it is more urgent to seek reliable audio tampering detection methods. Digital audio tampering detection based on power network frequency (ElectricNetwork Frequency, ENF) is a kind of evidence gathering method which has been paid more attention in recent years. However, the existing methods based on ENF have some shortcomings, such as the need of ENF reference database in judging tampering. Tampering with positioning accuracy is not high. Aiming at these shortcomings, the blind detection method of audio tampering without ENF reference database is studied in this paper. The main work and innovation are as follows: (1) A maximum correlation offset (Maximum Offset for CrossCorrelation, using ENF signal is proposed. MOCC). The ENF signal is divided into sub-blocks and the offset of the initial sub-block and each sub-block is calculated when the maximum cross-correlation value is obtained. The audio tampering is judged according to the consistency of each sub-block MOCC, and the tamper region location and tamper type estimation are realized. For different application scenarios, visual detection, automatic detection and fast detection are also proposed. Experiments show that the method can locate tamper accurately and judge whether the tamper type is deleted or inserted according to whether the content of the tamper region is silent part or speech part. In addition, because the method is detected in the time domain and does not need to be converted to the transform domain, the computational complexity of the method is smaller than that of the existing methods based on ENF phase. (2) aiming at the problem that the method in (1) is disturbed by noise when calculating MOCC, an improved MOCC audio tamper detection method is proposed. An ideal sine wave is introduced as the reference signal to enhance the ENF signal by using the multiple cross-correlation between the reference signal and the ENF sub-block signal, and the MOCC, of the enhanced ENF sub-block signal is calculated by using the reference signal. According to the change of MOCC, the tamper is located and the type of tamper is judged. Experiments show that this method can effectively improve the signal-to-noise ratio of ENF signal, reduce the interference of MOCC, and improve the accuracy of tampering detection under low false alarm rate. In addition, there are some robustness in resisting additive noise, audio resampling and audio compression. (3) A dual-mechanism audio tamper detection method based on minimum average amplitude offset (Minimum Offset for AverageMagnitude Difference, MOAMD) is proposed to improve the accuracy of tamper location. The audio tamper is detected by MOAMD to simplify the calculation of MOCC, and the tamper region is located by combining the curve variation of MOAMD and the slope of extreme point of MOAMD curve. Experiments show that this method can effectively improve the accuracy of tampering localization, and the two-mechanism can be extended to other methods. (4) an audio tamper detection method based on fast transversal filtering (Fast Transversal Filter,FTF (ENF neighborhood correlation coefficient) is proposed to improve the accuracy of existing methods. The correlation coefficient of the neighborhood block of ENF is calculated, and the correlation coefficient is filtered by FTF adaptive filter. The tampering is judged according to the change of error energy after filtering, and the tamper location is realized. Experiments show that this method can effectively improve the accuracy of tamper detection, and its advantages are more obvious in the case of large fluctuation range of ENF and low signal-to-noise ratio (SNR).
【学位授予单位】：华南理工大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN919.8

【相似文献】