基于期望最大化算法的音频取证中的篡改检测  

Audio re-sampling detection in audio forensics based on EM algorithm

作　　者：姚秋明[1] 柴佩琪[1] 宣国荣[1] 杨志强[1] 施云庆 

机构地区：[1]同济大学计算机科学与技术系,上海200092 [2]新泽西理工学院电气与计算机工程系,美国新泽西nj07102

出　　处：《计算机应用》2006年第11期2598-2601,共4页journal of Computer Applications

基　　金：国家自然科学基金资助项目(90304017)

摘　　要：音频取证中的插值检测是信号篡改检测的重要方面。因为信号的篡改经常伴随着重采样操作,而重采样后的插值信号会引入周期性信息。应用期望最大化(EM)算法能针对这种周期信息估计参数,从而检测出信号是否被篡改。为了使EM算法迭代效果更好,更适用于音频信号的插值检测问题,提出针对音频信号的特点,引入音频幅度直方图,排除短时静音和增加样本点数的方法。另外还提出了用频谱统计矩作为特征的方法,使统计分类稳定有效。最后通过音频取证中检测信号是否重采样的统计分类实验,表明整个检测流程能达到较高的准确率,并且在局部篡改实验中也同样有效。In audio forensics, re-sampling detection is an important part of signal interpolation detection, for the reason that audio interpolation is often accompanied with re-sampling. Re-sampled signals usually contain some periodic artifacts, which can be a proof of forgery. EM algorithm was applied in parameter estimation of the periodic artifacts to detect if the signal was interpolated. To work out a satisfying result in applying EM algorithm and make it more suitable to audio signal detection, audio signal histogram was introduced to describe the exact distribution, the silence and the local linear relationship was taken into account. In addition, the statistical moments in frequency domain were chosen as features to classify the original and the interpolated audio signals. The whole detection process was used on statistical classifying experiments. The result shows that it can get a high correct rate. It also took effect in local re-sampling detection.

关 键 词：音频取证 重采样检测 期望最大化算法 统计矩 

分 类 号：TP309.2[自动化与计算机技术—计算机系统结构]