上海交通大学 硕士学位论文 JPEG图像隐写信息盲检测系统的设计与开发 姓名：茹炳晟 申请学位级别：硕士 专业：软件工程 指导教师：黄林鹏;李华山 20081229 JPEG 图像隐写信息盲检测系统的设计与开发 I JPEG 图像隐写信息盲检测系统的设计与开发 摘 JPEG 图像隐写信息盲检测系统的设计与开发 摘 要 要 隐写术与隐写分析是互联网时代信息战一项重要内容， 对于信息安全具有重要意 义。目前，隐写术被非法组织用于传递敌对信息的可能性引起了广泛关注，同时，隐 写分析也受到了各方面的高度重视。另外，隐写术在反盗版领域也有广泛的应用。基 于以上情况，结合作者所在公司一个大型对日开发项目的实际要求，拟定设计开发一 个高性能的，可扩展的，具有较高识别率，并且具有较大实际应用价值的 JPEG 图像 隐写信息盲检测系统。 高性能是指所要开发的系统在对待检图片是否含有隐写信息进行检测时， 在保证 较高正确识别率的前提下，具有较好的时间效率，即要求在尽可能少的时间内完成对 图片的正确检测。可扩展性是指系统可以应对新出现的图像隐写工具，也就是说，对 于用新出现的图像隐写工具嵌入过信息的非自然图片，只要提供相应的训练样本集， 产生对应的分类器数据，本系统就能够对该类非自然图片具有较高的正确识别率。同 时还要求系统的分类器数据可以不断累积完善并有效共享，即要求分类器数据的使 用、保存、共享、新增以及补充完善具有灵活性，以此提高系统的实际应用价值。 基于以上要求，本文首先在综合考虑识别率和性能的基础上，提出了一种基于小 波变换，高阶统计量以及 Fisher 准则线性分类器的 JPEG 图像隐写信息检测算法，该 算法首先对图像进行二层小波分解，然后提取高阶统计量，得到描述图像统计特征的 特征向量，最后采用模式识别中的经典分类方法-2 类别 Fisher 准则线性分类器对数 字图像是否含有隐藏信息进行判决分类。期间对算法中的关键实现步骤，包括图像的 小波分解和 Fisher 准则线性分类器，编写了详细的类 C 伪代码，完成了系统核心算 法原型模块的开发。 接下来，在此算法设计的基础上，基于分类器数据模块化的设计思想并结合数据 JPEG 图像隐写信息盲检测系统的设计与开发 II 库技术，提出了一种全新的 JPEG 图像隐写信息盲检测系统的设计方案，基于该方案 开发的系统不仅具有良好可扩展性，以应对新出现的图像隐写工具，还能不断累积完 善并有效共享分类器数据，由此可以延长系统的生命周期。并且系统还能在使用过程 中不断提高检测识别率，具有较大的实际应用价值。 最后，基于 Jsteg 大型训练图像样本集的仿真试验取得了预计的结果，验证了系 统设计的合理有效性，很好满足了公司项目对系统的设计要求。 关键词 关键词 信息隐藏，盲检测，FLD，可扩展性，高阶统计量，小波分解 JPEG 图像隐写信息盲检测系统的设计与开发 III DESIGN AND DEVELOP OF BLIND STEGANALYSIS SYSTEM BASED ON JPEG IMAGES ABSTRACT Steganography and steganalysis, which are very important component of information warfare in the age of the Internet, have great significance for information security. At present, the possibility of using steganography for hostile information transmission by illegal organizations has caused wide public concern. Meanwhile steganalysis has been attached great importance from all sides. Besides, steganography also finds wide application in the field of anti-piracy. Based on the above, we prepared to design and develop a new Blind Steganalysis System Based on JPEG Images with high-performance, extensible, high recognition rate, and greater practical value, which are according to actual requirements of a large-scale Japanese project in authors company. High-performance means a high rate of correct identification with great time efficiency when the system is used to decide whether images contain any hidden information. Thats to say, the system should complete the detection correctly as soon as possible. Extensibility refers to that the system can deal with new steganography tools for JPEG images, which is to say, if we provide training sample set with security messages embedded by new steganography tool and establish the corresponding classifier data, the system can identify these non-natural images exactly. Also, the system requires that the classifier data can be continuously accumulated and effectively shared. The usage, preservation, sharing, as well as accumulation of classifier data should obtain flexibility to increase the systems practical value. Based on the above requirements and an overall consideration of recognition rate and time efficiency, this paper presents a arithmetics of JPEG images blind steganalysis using wavelet decomposition, higher-order statistics and Fisher linear discriminant. In this arithmetic, a 2-level wavelet decomposition is used on images and images are characterized based on building higher-order statistical model to obtain feature vector of each image. At last, the two-class Fisher linear discriminant analysis (FLD), a class specific method for pattern recognition, is used to decide whether images contain any JPEG 图像隐写信息盲检测系统的设计与开发 IV hidden message or not. The paper also implents the key steps of prototype algorithm module with C pseudocode, including wavelet decomposition on images and Fisher linear discrimination. According to above arithmetics, this paper proposes a new design scheme of Blind Steganalysis System Based on JPEG Images based on modularization design of classifier data and database technology. The system based on this design scheme not only has good extensibility to handle the new steganography tools for JPEG image, but also be able to accumulate and share the classifier data continuously and effectively, which can extend the systems life cycle. Besides, the system can continuously improve recognition rate and has great practical value. Result of simulational experiment on Jstegs large-scale training set shows that the system design has achieved expected results with verified rationality and availability. So it well satisfies the design requirements of companys project. Keywords hidden information, blind steganalysis, FLD, extensible, higher-Order Statistics, wavelet decomposition JPEG 图像隐写信息盲检测系统的设计与开发 上海交通大学上海交通大学 学位论文原创性声明学位论文原创性声明 本人郑重声明：所呈交的学位论文，是本人在导师的指导下，独立进行 研究工作所取得的成果。除文中已经注明引用的内容外，本论文不包含任 何其他个人或集体已经发表或撰写过的作品成果。对本文的研究做出重要 贡献的个人和集体，均已在文中以明确方式标明。本人完全意识到本声明 的法律结果由本人承担。 学位论文作者签