基于机器学习的AbS-LPC压缩语音隐写分析研究

项目名称： 基于机器学习的AbS-LPC压缩语音隐写分析研究

项目编号： No.61303249

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 李松斌

作者单位： 中国科学院声学研究所

项目金额： 23万元

中文摘要： 现有的IP网络语音通信系统，多采用按合成分析-线性预测编码（AbS-LPC）方法对语音进行压缩编码，这就在网络中产生了大容量且具有瞬态性的AbS-LPC压缩语音码流--极具威胁性的信息隐藏载体。AbS-LPC码流中的信息隐藏是在压缩域进行的，对压缩语音引入的附加失真极小，因此很难获得辨识是否隐写的特征，隐写检测难度极大。为解决这一问题，本课题拟基于机器学习框架对AbS-LPC码流中的三类典型信息隐藏方法展开研究：（1）拟基于语音产生模型及语音学中的音素分布特性，探索多矢量联合MEQIM信息隐藏的检测方法；（2）拟从浊音音素所覆盖的相邻语音帧基音的共生特性出发，探索基音周期调制信息隐藏的检测方法；（3）拟基于AbS-LPC压缩编码无法去除语音信号噪声的局部相关性这一发现，研制一种较具普适性的码元LSB替换隐写检测算法。在此基础上，设计并开发一个AbS-LPC压缩语音信息隐藏综合检测平台。

中文关键词： 合成分析线性预测编码；压缩语音；隐写分析；特征抽取；机器学习

英文摘要： In existing voice over IP communicaiton systems, the Analysis-by-Synthesis Linear Predictive Coding (AbS-LPC) method is widely used for speech compression coding, which makes AbS-LPC compressed speech stream an extremely threatened carrier for information hiding. Because they hide the secret bits in the compressed-domain, the additonal distortion of speech introduced by the information hiding methods based on AbS-LPC compressed speech is very minimal. Therefore, it is very difficult to acquire the feature that can be used for distinguishing whether informaition hiding exists in AbS-LPC compressed speech. How to detect this kind of information hiding effectively becomes a challenging problem. This proposal aims to solve this problem through studying on steganalysis of three kinds of typical information hiding methods of AbS-LPC compressed speech, under the machine learning framework. Firstly, this proposal will investigate the detection method of multi-vector association MEQIM stegangraphy, according to the speech generation theory and the phoneme distribution characteristics of Phonetics. Secondly, starting from the co-occurrence properties of pitches of adjacent frames covered by voiced phonemes, this proposal will explore the steganalysis algorithm for pitch-period-modulation steganography. Thirdly, based on o

英文关键词： AbS-LPC；Compressed Speech；Steganalysis；Feature Extraction；Machine Learning