目的 隐写分析研究现状表明，与秘密信息的嵌入过程相比，图像内容和统计特性差异对隐写检测特征分布会造成更大的影响，这导致图像隐写分析成为了一个“相同类内特征分布分散、不同类间特征混淆严重”的分类问题。针对此问题，本文提出了一种更加有效的JPEG图像隐写检测模型。方法 通过对隐写检测常用的分类器进行分析，从降低隐写检测特征类内离散度的角度入手，将基于图像内容复杂度的预分类和图像分割相结合，根据图像内容复杂度对图像进行分类、分割，然后分别对每一类子图像提取高维富模型隐写检测特征，构建分类器进行训练和测试，并通过加权融合得到最终的检测结果。结果 在实验部分，对具有代表性的隐写检测特征集提取了两类可分性判据，对本文算法的各类别、区域所提取特征的可分性均得到明显提高，证明了模型的有效性。同时在训练、测试图像库匹配和不匹配的情况下，对算法进行了二分类测试，并与其他算法进行了性能比较，本文方法的检测性能均有所提高，性能提升最高接近10%。结论 本算法能够有效提高隐写检测性能，尤其是在训练、测试图像库统计特性不匹配的情况下，本文算法性能提升更加明显，更适合于实际复杂网络下的应用。
Abstract ： Objective Image steganalysis is the opposite technology against steganography, which aims at detecting, extracting, restoring and destroying the secret message embedded into the cover images. As a quite important technical tool for image information security, image steganalysis has become an attractive hotspot of the multimedia information security to researchers all over the world. The basic concept of the current image steganalysis is to analyze the embedding mechanism and the statistical changes of the image data caused by embedding the secret messages. And the images steganalysis is dealt as a binary classification problem, in which the cover images and the stego images are seemed as two categories to be classified. The performance of the steganalysis methods depends on feature extracting, and the steganalysis features are expected to have small within-class scatter distances and big between-class scatter distances. However, the embedding changes are not only correlated with the steganography methods, but also with the image content and the local statistical characteristics. The changes of the steganalysis features caused by secret embedding are subtle, especially when the embedding ratio is low. Compared with the process of embedding, the contents and statistical characteristics of the images make a much stronger impact on the distribution of steganalysis features. As a result, the steganalysis features of the cover and stego images will be inseparate caused by the differences of image statistical characteristics. This makes the image steganalysis become a classification problem with bigger within-class scatter and smaller between-class scatter distances. To solve this problem, a new steganalysis framework for JPEG images which aims at reducing the within-class scatter distances was proposed. Method After embedding, the secret messages will have different effects on the characteristics of images with different content complexities, whereas the steganalysis features of the images with the same content complexity are similar. This paper focuses on image steganalysis based on reducing the differences of image statistical characteristics caused by various content and processing methods. In this paper, the motivation of the new model was introduced by analyzing the Fisher Linear discriminant analysis, which is the basic of ensemble classifier, the most used one in steganalyzing application, and a new steganalysis model of JPEG images based on image classification and segmentation was proposed. We define a content complexity evaluation feature to each image, and the given images were first classified according to the content. As a result, the images classified to the same sub-class will have closer content complexity. Then each image was segmented to several sub-images according to the texture features evaluated the complexity of each sub-block. When segmenting, we first categorize the image blocks according to the texture complexity, and then amalgamate the adjacent block categories. After the classification and segmentation combined processing, the content texture of the same class of image regions will be much similar, and the steganalysis features will be more centralized. The steganalysis features were separately extracted from each subset with the same or close texture complexity to build a classifier. When deciding which steganalysis feature set to be extracted, we considered the performance mainly. In our prior work, we find that when extracting steganalysis feature set with low dimension, the performance of the method based on classification or segmentation can be obviously improved. However, when extracting high dimensional steganalysis features, such as JPEG rich model (JRM), the performance was not so satisfactory, which is because the rich model is based on residual of the given image, and it can eliminate the effect of image content. JRM feature set is sensitive to subtle image details, and the steganalysis result is good enough. However, in this paper, we still extract JRM feature set, which is the most representative high dimensional feature set in JPEG domain, to prove the validity of the proposed model. In the testing phase, the steganalysis features of each segmented sub-image in each sub-class are sent to the corresponding classifier. The final steganalysis result was figured out through a weighted fusing process. Result In the experiment, we compute two kinds of separability criterion of the tested steganalysis feature set, including separability criterion based on within and between class distances and the Bhattacharyya distances. The Bhattacharyya distance is one of the most used separability criterion on the basis of the probability density of the classified samples. Both separability criterion of the proposed method are obviously improved, which means the proposed classified and segmented based steganalysis features can be more easily categorized. This can verify the validity of the proposed steganalysis model. We also compare the classified performance of the proposed method and the prior work in various experimental circumstances, including the training and testing image database are the same and different. We compute the detecting result of the original feature set, the features extracted from the classified image, segmented image, and the classification and segmentation combined image. The experimental result show that in both the two circumstances, classification and segmentation combined processing can effectively improve the performance, and the improvement can be up to 10%. The improvement is much considerable when training and testing images are with different statistical features, which implies that the proposed method is more suitable for practical application for the images on the Internet are with considerable diversities in sources, processing methods and contents. Conclusion In this paper, a new steganalysis model to JPEG images were proposed. The differences of image statistical characteristics caused by various content and processing methods were reduced by image classification and segmentation. JRM feature set was extracted. The theoretical analysis and experimental results on several diverse image databases and circumstances can demonstrate the validity of the framework. When there is a considerable diversity in image sources and contents, such as the training and testing image are different, the performance improvement of the proposed method are even more obviously, which indicates that the performance of the proposed method are not highly depending on the image content. It implies that the proposed steganalysis model is more suitable for practical application in complex network environment.