基于统计模型及SVM的低速率语音编码QIM隐写检测  

Detection of QIM Steganography in Low Bit-Rate Speech Codec Based on Statistical Models and SVM

作　　者：李松斌[1,2] 黄永峰[1] 卢记仓[3] 

机构地区：[1]清华大学电子工程系,北京100084 [2]中国科学院声学研究所南海研究站,海口570105 [3]解放军信息工程大学网络工程系,郑州450002

出　　处：《计算机学报》2013年第6期1168-1176,共9页Chinese Journal of Computers

基　　金：国家自然科学基金(60970148,61271392)资助~~

摘　　要：QIM(Quantization Index Modulation,量化索引调制)隐写在标量或矢量量化时嵌入机密信息,可在语音压缩编码过程中进行高隐蔽性的信息隐藏,文中试图对该种隐写进行检测.文中发现该种隐写将导致压缩语音流中的音素分布特性发生改变,提出了音素向量空间模型和音素状态转移模型对音素分布特性进行了量化表示.基于所得量化特征并结合SVM(Support Vector Machine,支持向量机)构建了隐写检测器.针对典型的低速率语音编码标准G.729以及G.723.1的实验表明,文中方法性能远优于现有检测方法,实现了对QIM隐写的快速准确检测.Quantization Index Modulation （QIM） steganography, which embeds the secret infor- mation during the Vector Quantization, can hide information in low bit-rate speech codec with high imperceptibility. This paper tries to detect this type of steganography. For this purpose, starting from the speech generation and compress coding theory, this paper firstly analyzes the possible significant feature degradation through the QIM steganography in compressed audio stream deeply. And it finds that the QIM steganography will disturb the phoneme sequence in the stream, and inevitably make the imbalance and correlation characteristics of phoneme distribution in the sequence change. According to this discovery, this paper adopts the phoneme distribution characteristics as the key for the detection of the QIM steganography. In order to get the quanti- tative features of phoneme distribution characteristics, this paper designs the Phoneme Vector Space Model and the Phoneme State Transition Model to quantify the imbalance and correlation characteristics respectively. By combining the quantitative vector features with supervised learn- ing classifier, this paper builds a high performance detector towards the QIM steganography in low bit-rate speech codec. The experiments show that, for the two typical low bit-rate speech codec： G. 729 and G. 723.1, the proposed method has an excellent performance compared to existing method.

关 键 词：QIM隐写 隐写检测 低速率语音编码器 音素分布特性 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]