【作者】 肖博；

【导师】 黄永峰；

【作者基本信息】 清华大学 ， 信息和通信工程， 2009， 硕士

【摘要】 低速率语音编码中的信息隐藏问题是信息隐藏领域的难点,也是构建信息隐藏通信系统的基础。通过修改编码的矢量量化环节实现嵌入是一种主要方法,其中码本分组是否合理决定了信息隐藏的隐蔽性优劣。本文基于QIM方法和图论,提出了优化的码本分组算法,即互补邻居顶点算法(CNV)。该算法对于任给的码本,可以保证划分结果使得每个码字和它最近邻的码字分属不同的分组,且使得局部附加量化失真的极大值对各种划分方式取得其极小。在理论方面,本文首先利用图论建模,证明了上述两个结论。接着,通过反例分析了满足约束使得每个码字和它最近的两个或三个邻居顶点两两属于不同的分组的分组方式对任意码本不能普遍成立。为了分析给定的分组是否存在满足上述条件的分组方式,我们提出了基于回溯法的搜索方法。此外,提出了重复使用CNV算法将分两组的结果各自再分一次,得到近似优化的分四组的方法,改进了信息隐藏的容量。在实际编码方面,本文使用CNV算法对iLBC、G.729、G.723.1三种编码的LPC系数矢量量化码本做了划分,给出了分组结果并做了讨论。回溯搜索的结果表明不能对上述码本按照类似CNV的约束条件分多组,但近似优化的分四组的方法可以很好的成立。通过大量实验,发现CNV算法划分的平均量化误差小于随机划分,且有利于取得较小的最大量化误差。将上述划分应用于实际信息隐藏,对语音质量的主观评价表明主观感受不能区分嵌入机密信息的合成语音与普通的合成语音。采用平均LPC倒谱失真作为客观评价准则,发现CNV算法划分具有较好的语音质量,而近似分四组方法的效果也好于直接替换量化结果这一类方法。最后,将CNV方法应用于前期工作基于VoIP的信息隐藏通信系统中,通过实际应用验证了其有效性,传输速率和理论分析一致,满足实际应用需求,是目前较理想的隐藏方法。更多还原

【Abstract】 Information Hiding (IH) in low bit-rate speech codec is a di?cult problem in IHresearch as well as the foundation of building IH communication system. One mainembedding method is modifying Vector Quantization (VQ) procedure during speechencoding, where the covertness of IH is determined by codebook partition. Basedon the QIM method and Graph Theory, the Complementary Neighbor Vertex (CNV)method is proposed in this paper. CNV method guarantees that for any given code-book,every codeword belongs to the di?erent part of its nearest neighbor’s. It alsoensures that the maximum of local extra quantization distortion due to IH reaches itsminimum among all dividing patterns.Theoretically, This paper first establishes analytical model using Graph Theory,and gives proof to the above conclusions. Secondly, it is demonstrated using counterexample that CNV-like dividing, that is, making every codeword and it’s two or threenearest neighbor each having di?erent color, is not generally feasible towards any code-book. In order to find out whether such dividing pattern exists, a searching algorithmbased on backtracking is proposed. In addition, another high capacity algorithm is in-troduced using CNV method iteratively, which divides the 2 parts derived from CNVmethod again into 4 parts approximately optimized.Practically, The CNV method is used in dividing the VQ codebooks of LPC co-e?cients in iLBC, G.729, and G.723.1 codec. The result of dividing is given anddiscussed. The attempt to divide the codebooks into multiple parts with CNV-likerestrictions is not successful. Nevertheless, dividing those codebooks into 4 parts ap-proximately optimized is proved applicable.Extensive experiments in real audio resources demonstrate that the mean quanti-zation error of CNV method was smaller than that of dividing the codebook randomly,and it is more likely to get a smaller maximum quantization error using CNV method.Subjective quality assessment of the produced audio signal shows that it is impossi- ble to distinguish the audio signal containing embedded messages from normal recon-structed audio signal by human ears. Adopting the Mean LPC Cepstrum Distortion asobjective evaluation method, it is found that using CNV method is beneficial for get-ting better quality. And dividing the codebook into 4 parts approximately optimizedexceeds the class of methods which replace the quantization result directly after encod-ing.Finally, the CNV method is applied to our previous work which was a covertcommunication system based on VoIP and IH. Though practical test, the feasibility ofour methods is proved, and the measured transmission rate is consistent with theoreticalanalysis. The CNV method satisfies application requirement, being the favorable IHmethod in the system up to now.更多还原