数字图像隐写分析研究

发布时间：2018-01-08 20:25

本文关键词：数字图像隐写分析研究　出处：《上海大学》2016年博士论文　论文类型：学位论文

【摘要】：数字图像隐写和隐写分析技术是两个对立的学科。隐写技术主要利用多媒体数据实现信息的秘密传递,即隐蔽通信。隐写分析是发现隐写行为的技术。本文以隐写分析技术为研究目标,结合内容自适应隐写方法,提出了多种有效的隐写分析方法。同时,为了提高分类器的检测能力,结合集成学习理论,提出了多样化集成分类器。取得的成果如下:1.基于重分布残差的低维空域图像隐写分析提出了2929维重分布残差空域图像隐写分析特征。该特征由两类子特征构成,第一类子特征是残差的一阶直方图。该部分特征由较大阈值门限生成,能够捕捉分布在边缘以及纹理区域的隐写改变。第二类子特征是重分布残差的一阶直方图。该部分特征利用平移参数和残差的反相关性生成,重分布机制能够增加特征的多样性,提高特征的检测准确率。为了检测特征的有效性,我们选用内容自适应安全隐写算法进行测试。实验结果显示,在低嵌入率下,同已有的隐写分析方法相比,重分布特征对于内容自适应隐写算法的检测错误率降低了5.65%。2.基于纹理复杂度和残差对比度的低维空域图像隐写分析提出了基于纹理复杂度和残差对比度的2363维空域图像隐写分析算法。根据内容自适应隐写算法原理,秘密信息被嵌入到图像的纹理以及边缘区域,而图像的平滑区域不会嵌入秘密信息。因此,为了能够根据图像的纹理进行针对性特征提取,定义了波动函数评价图像的纹理复杂度,然后根据波动函数提取出图像中最复杂的区域(子图像)。对于提取的子图像和原始图像,算法使用线性及非线性滤波器获得多样化残差,然后将不同残差的比值转换成角度。同时,将残差的l2范数作为角度的对应权值,最终将加权的角度直方图作为残差对比度特征。该特征不仅可以有效的表达不同残差之间的联合统计分布,而且特征维数随着阈值门限值呈线性变化。同已有的空域图像隐写分析算法相比,在较低的嵌入率下,提出的隐写分析算法使用低维特征可以实现较高的检测准确率。3.基于多分辨率分解和仿射变换的JPEG图像隐写分析提出了基于多分辨率分解和仿射变换的高维JPEG图像隐写分析算法。我们认为JPEG图像是由若干不同分辨率的子图像以非线性方式构成。因此,如果可以将原始JPEG图像进行多分辨率分解,那么隐写改变会以非线性方式分布于不同的JPEG子图像中。同时,通过逐步剥离图像的平滑区域,可以更好的凸显出隐写改变,从而增加特征的检测准确率。为了获得多分辨率的JPEG图像,首先将JPEG图像解压缩成空域图像,在空域中利用多分辨算法进行分解。然后将不同分辨率的空域子图像重新压缩,获得多幅JPEG图像。其次将每个DCT(离散余弦变换)系数看成是独立的平面,按照不同方向获取残差平面。将两个残差系数间的比值转换成角度,同时将DCT系数的l2范数作为角度的权重,最终把角度和权值的联合作为提取的特征。为了增加特征的多样性,我们使用仿射变换对转换的角度进行旋转,从而获得新的特征。相比已有JPEG图像隐写分析算法,在不同嵌入率下,提出的隐写分析算法能够提高特征的检测准确率。4.多样化集成分类器提出了多样化集成分类器算法。集成分类器能够解决高维特征的训练与分类,但是在原始分类器中仍然存在两个不足。第一个不足是最终分类器的选择策略。面对众多训练好的基分类器,原始的算法会丢弃绝大多数的基分类器,只选择具有最小训练误差的基分类器作为最终分类器。为了防止过拟合现象发生,增强最终分类器的泛化能力,我们对全部基分类器使用Bagging集成策略,生成最终分类器。该算法能够充分利用被丢弃的基分类器,从而增强其泛化能力,避免过拟合现象。第二个不足是子分类器的弱分类能力会影响最终分类器的检测性能。为了提升子分类器的分类性能,我们采用Adaboost集成学习策略提升单一子分类器的分类能力,最终提升分类器的整体检测能力。对于多种隐写分析特征的检测结果显示,利用多样化集成分类器可以提高特征的检测准确率。本文以隐写分析技术为目标,从重分布残差构造、纹理复杂度和残差对比度特征提取、多分辨率分解和多样化集成分类器设计等角度对隐写分析技术进行了分析与研究。
[Abstract]:Digital image steganography and steganalysis are two opposite subjects. Steganography mainly use multimedia data to realize the information transmission of secret covert communication. Namely, steganalysis is found in steganography technology. Based on the behavior of steganalysis technique as the research object, combined with the content of adaptive steganography method is proposed a variety of effective steganalysis methods. At the same time, in order to improve the detection capability of classifier, integrated learning theory, put forward the diversified integrated classifier. The results are as follows: 1. based on the low dimensional spatial domain image hidden redistribution residual analysis proposed 2929 dimensional spatial redistribution of residual image steganalysis features. The characteristics of composed of two kinds of features, the first sub feature is a first-order histogram. The residual part features generated by the larger the threshold, to capture changes in the distribution of write edges and texture regions. Second kinds of hidden Feature is the first-order histogram redistribution of residual. This part features using the translation parameters and residual anti correlation generation, redistribution mechanism can increase the diversity characteristics, improve the feature detection accuracy. In order to check the validity of the feature, we use content adaptive security steganographic algorithm is tested. Experimental results show that in low the embedding rate, with the existing steganalysis method compared to heavy distribution for the detection of the error of adaptive steganography algorithm based on 5.65%.2. to reduce the rate of low dimensional spatial images texture complexity and residual contrast analysis proposed 2363 dimensional spatial domain image texture complexity and implicit residual contrast analysis algorithm based on according to the content of adaptive steganography algorithm principle, the secret information is embedded into the image texture and edge region, and the secret information of the image smoothing area not embedded. Therefore, for It can be targeted according to the image texture feature extraction, texture image definition evaluation function fluctuation complexity, then according to the wave function to extract the most complex regions in the image (sub images). The extracted sub image and the original image. The algorithm uses a linear and nonlinear filter to obtain diverse residuals, then different residuals the ratio of conversion into perspective. At the same time, the L2 norm residuals as corresponding weight angle, will eventually be weighted angle histogram as residual contrast features. This feature not only can be expressed efficiently between different residual joint statistical distribution, and the threshold value dimension with linearly implicit spatial image with the existing writing the analysis algorithm, at low embedding rate, the proposed steganalysis algorithm using low dimensional feature can achieve higher detection accuracy of.3. based on multi resolution JPEG images decomposition rate and affine transform writing analysis put forward high dimensional JPEG image hidden multi-resolution decomposition and affine transform analysis based algorithm. We believe that the JPEG image is composed of a plurality of different resolution sub image in a nonlinear way. Therefore, if the original JPEG image multi-resolution decomposition, then steganography change be in a nonlinear way distributed in different JPEG sub image. At the same time, the smooth region gradually stripped of the image, you can better highlight the steganography change, thereby increasing the accuracy of the feature detection. In order to obtain the JPEG image resolution, the JPEG image decompression into spatial domain image, using multiresolution decomposition algorithm in the spatial domain. Then different spatial resolution image compression subsystem, to obtain JPEG images. Then each DCT (discrete cosine transform) coefficients as independent The plane, according to the different direction. The plane gets residual ratio of two residual coefficient of conversion between angle, while the L2 norm DCT coefficients as angle weights, finally put joint angle and weight as the extracted features. In order to increase the diversity characteristics, we use affine transform to convert rotation angle in order to obtain new features. Compared with the existing JPEG image steganalysis algorithm in different embedding rate, the proposed steganalysis algorithm can improve the detection accuracy of.4. features of diverse ensemble classifier proposed diverse ensemble classification algorithm for training and classification. The ensemble classifier can solve high dimensional features, but still in the original classifier there are two problems. The first problem is the final classifier selection strategy. In the face of many base classifiers trained, the original algorithm will discard most of the base class Is only the base classifier with minimum training error of the final classifier. In order to prevent overfitting phenomenon, finally enhance the generalization ability of the classifier, we have all the base classifiers using Bagging integration strategy, generate the final classifier. This algorithm can make full use of the base classifier is discarded, so as to enhance its generalization ability, avoid the overfitting phenomenon second is the lack of detection performance of weak classification ability of classifier will affect the final classifier. In order to improve the classification performance of classifier, we use the classification ability of Adaboost integrated learning strategy to improve single classifier, improve the overall detection ability of the classifier. The final results for the detection of a variety of steganalysis features, using diverse ensemble classifier can improve the feature detection accuracy. Based on the steganalysis technology as the goal, from the distribution of residual. The steganalysis technology is analyzed and researched based on texture complexity and residual contrast feature extraction, multi-resolution decomposition and diversified ensemble classifier design.

【学位授予单位】：上海大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41;TP309

【相似文献】