面向CELP语音压缩域的通用隐写分析方法研究

项目名称： 面向CELP语音压缩域的通用隐写分析方法研究

项目编号： No.U1536114

项目类型： 联合基金项目

立项/批准年度： 2016

项目学科： 管理科学

项目作者： 任延珍

作者单位： 武汉大学

项目金额： 64万元

中文摘要： 码激励线性预测（CELP）是目前应用最广泛的语音编码模型，其压缩域隐写算法已出现，隐藏容量大且隐蔽性好，给隐写分析带来新的挑战。本课题研究CELP压缩域通用隐写分析方法。在特征提取方面，针对固定码本、基音周期和LPC系数嵌入域，提出基于固定码本最优搜索原则、基于基音周期多阶相关性度量、以及基于距离的LPC码本统计分布特性的隐写分析特征，提高特征的通用性；在隐写分析学习方法方面，针对失配问题所带来的检测率下降问题，提出基于样本层和特征表示层域自适应学习的隐写分析方法，通过源域和目标域的映射和模型迁移，提升算法的检测性能；针对未知算法的检测，提出基于半监督学习的隐写分析方法，通过基于聚类的选择性集成学习训练协同分类器，提升算法的泛化能力。课题研究成果可应用于3G通信、移动互联网以及VoIP通信中，实现对CELP压缩语音的隐写检测。课题隐写分析学习方法可为图像视频隐写分析提供新思路。

中文关键词： 隐写；隐写分析；CELP；语音；压缩域

英文摘要： Code Excited Linear Prediction (CELP) is the most popular speech coding model, the steganography algorithms based on CELP compressed domain have been emerging and are quite promising for both high capacity and good imperceptivity, which bring new challenges to steganalysis. The universal steganalysis method on CELP compression is studied in the project. To extract steganalysis features for FCB (Fixed Codebook), pitch delay and LPC coefficients, we proposed the steganalysis feature based on the optimal FCB search principle, multi-order correlation measure of pitch period and the statistical distribution characteristics based on distance of LPC codebook to improve the feature universality. To solve the problem of mismatch, domain adaption method is proposed, from sample domain and feature domain respectively, remodeling and mapping the distribution of source and target domains to improve the steganalyer’s efficiency. To improve the detection performance for unknown steganagraphy methods, steganalysis based on Semi-Supervised Learning are put forward, we adopt collaborative classifier trained by selective ensemble learning based on clustering to improve the generalization ability of the algorithm. Research results can be applied to 3G communications, mobile Internet and VoIP to realize the steganalysis detection of CELP compressed speech. The steganalysis learning methods can provide new methods for image and video steganalysis.

英文关键词： steganagraphy;steganalysis;CELP;Audio;Compressed Domain