Abstract:

Digital audio source passive forensics research aims at extracting the characteristics of the machine fingerprint through the intrinsic device information obtained from the original digital audio data without relying on the redundant information such as digital watermark or digital signature. After that it can make a decision about the source of the digital audio evidence. It has broad application prospects in the fields of judicial forensics, military information, and news dissemination, etc. At present, the research review on passive forensics of digital audio sources faces the problem of insufficient timeliness and insufficient targeting. Based on this, the framework and basic research ideas of passive forensics of digital audio sources are given. Then a brief analysis of the commonly used data sets in this field is made. According to the research object of digital audio sources passive forensics, the research in the field are divided into two modules： characteristic expression and characterization modeling module, and then the five kinds of characteristics such as frequency domain features, cepstrum features, based on the Gaussian super vector information features, fusion features, deep features and five types of forensics models such as Gaussian mixture model, support vector machine, sparse representation classifier, other machine learning methods, deep learning decision model are compared and analyzed. Finally, the research status and existing problems in the field of passive forensics of digital audio sources are summarized and analyzed, and the future research direction is prospected.

Key words: digital audio forensics, source recording device forensics, passive forensics, digital audio source identification

摘要：

数字音频来源被动取证研究旨在不依赖主动嵌入的数字水印或数字签名等冗余信息，通过原始数字音频数据的内在设备信息提取出表征设备源机器指纹的特征，进而对数字音频证据来源做出判断，在司法取证、军事信息、新闻传播等领域有着广泛的应用前景。目前，数字音频来源被动取证的研究综述面临时效性不足、针对性不够的问题。据此，给出了数字音频来源被动取证的研究框架和基本思路。对该领域常用的数据集做了简要的分析。根据数字音频来源被动取证的研究对象，将领域内的研究分为特征表达和表征建模两大模块，对频域信息特征、倒谱特征、基于高斯超矢量信息的特征、融合特征、深度特征五类特征，高斯混合取证模型、支持向量机决策模型、稀疏表达分类器决策模型、其他机器学习决策模型、深度学习决策模型五类模型的性能进行了比较分析。总结分析了数字音频来源被动取证领域的研究现状和存在的问题，并对未来的研究方向进行了展望。

关键词: 数字音频取证, 设备源取证, 被动取证, 数字音频来源识别