图像新型自适应隐写检测的去失配及敏感特征获取

The modern adaptive steganography, which based on the framework “distortion function + coding”, is the developing tendency of current image steganography. Traditional steganalysis methods can hardly achieve effective detection for these steganography. For the steganalysis of adaptive steganography at present, there are some problems which will influence the detection performance: the use of high dimensional or rich model features increase complexity of the detection algorithm, and there also may be redundant feature components. The problem of detection mismatch will happen when the distributions of training samples are different from that of testing samples. This project focuses on the key issues of local area distortion estimation, mismatch correction, feature extraction source constructing and sensitive feature obtaining. According to the modification and distribution of image data, deal with regional division for image. Based on the analysis of distortion function, the modification distortion of image data and local area are estimated, and then, the features that could depict the distortion distribution are extracted for image pre-classification and steganalytic mismatch correction. According to the local area distortion estimation result, determine the regions with high modification ratio, and then, construct sensitive feature extraction source and extract corresponding features. For the extracted features, analyze the separability based on feature separability principle for feature optimization, and then, the sensitive features for steganalysis of adaptive steganography can be obtained. The research of this project is expected to make some breakthroughs in the aspects of local area distortion estimation and mismatch correction, feature extraction source construction and sensitive feature obtaining, which could provide theoretical basis and technical support for the reliable detection of adaptive steganography.

基于“失真函数+编码”框架的新型自适应隐写是当前图像隐写发展的一个趋势，传统隐写检测方法难以对其实施有效检测。现有检测中，特征维数很高但存在冗余分量，检测复杂度较高，训练和检测样本数据分布不一致将引起检测失配，这些都将影响隐写检测性能。本项目着重对图像新型自适应隐写检测中的区域失真估计及去失配、特征提取源构建及敏感特征获取等关键问题展开研究，拟根据隐写更改特点和图像数据分布特性进行区域划分，基于失真函数分析对数据及区域更改失真进行定量描述，提取刻画失真分布的统计特征，进而对图像进行预分类以去除检测失配；基于区域失真估计选择高概率更改区域，在此基础上，构建敏感特征提取源并提取特征，基于特征可鉴别性原则分析其可分性并进行优化，进而获取自适应隐写敏感检测特征。项目期望在区域更改失真估计及检测去失配、特征提取源构建及敏感特征获取等方面有所突破，为新型自适应隐写的可靠检测提供理论依据和技术支持。

已有针对图像新型自适应隐写的检测通常基于“特征提取->分类器训练->分类检测”框架实现。其中，特征提取要求能够捕获隐写更改对数据造成的敏感变化，分类器训练要求利用与待检测图像类型和统计特性相似的载体和隐密图像来训练模板，以实现更好地分类检测。然而，已有方法特征提取存在盲目性、特征维数存在冗余性、分类器训练存在失配性，从而影响自适应隐写的检测性能。. 本课题重点对图像新型自适应隐写检测中的区域更改失真分析及去失配、敏感特征提取源构建、隐写检测特征提取及优化等问题进行了研究。在区域更改失真估计方面，分别从滤波系数权重分配及隐写选择信道等方面，提出了相应的特征提取及隐写检测方法；在图像预分类及隐写检测去失配方面，提出了一种基于大规模图像预分类和特征选择相结合的自适应隐写检测方法和一种结合图像操作脆弱检测的隐写检测框架；在隐写检测特征优化方面，研究了多种特征可分性准则，提出了多种富模型特征选取方法；在自适应隐写检测特征提取及检测方法方面，提出了一种用于空域隐写检测的敏感特征融合提取框架与改进方法、一种基于深度隐写检测特征学习的检测方法及多种JPEG图像自适应隐写检测特征提取和检测方法；此外，还对典型隐写嵌入比率估计、信息嵌入长度定量分析、隐写算法识别、典型变换操作检测等进行了研究，提出了基于隐写更改分析的嵌入率估计算法、多种隐秘信息嵌入位置估计算法、一种基于像素相关性特征的隐写算法识别方法及图像伽马变换参数估计和相应的图像拼接检测方法等。. 基于部分结果，在“Science China: Information Sciences”、“Signal Processing”、“IEEE Access”、“Journal of Visual Communication and Image Representation”、“Multimedia Tools and Application”、“Journal of Electronic Imaging”等国内外重要刊物发表学术论文24篇，其中SCI期刊论文16篇。申请国家发明专利3项。协助培养毕业博士生3名，硕士生2名。课题负责人卢记仓获ACM中国郑州分会新星奖，2018年作为讲师被破格遴选为硕士研究生导师。