【作者】 郭艳卿；

【导师】 何德全； 孔祥维； 尤新刚；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2009， 博士

【摘要】 几千年来,如何确保信息在通信过程中的安全一直是人类面临的重要研究课题。随着网络及多媒体技术的飞速发展,以“隐匿秘密通信存在”为目的的信息隐藏重新受到了人们的广泛关注。在这种通信范式下“隐藏信息”和“检测信息”的隐密对抗问题也迅速成为学术界的研究热点。本文针对目前隐密对抗领域仍存在的诸多关键问题,分别从隐密安全性、隐密方法和隐密分析方法三个方面对隐密对抗所涉及的理论及方法进行了系统研究,主要工作和成果总结如下:在隐密安全性的理论方面,首先分析了现有隐密安全性理论模型的局限性,阐明了区分“含密数据安全性”和“隐密算法(或系统)安全性”的重要意义,并在全信息视角下重新探讨了隐密安全性的内涵,指出隐密本质上源于认识主体对客观事物全信息的有限认识能力。在此基础上引入“载体分布边界”的概念,构建了新的统计安全性理论框架,明确给出了含密数据安全性、隐密算法安全性和载体数据安全隐密容量的定义,并建立了隐密算法“未公开”和“已公开”两种情况下的隐密安全性描述及度量模型。在隐密模型及方法的设计方面,首先分析了典型隐密技术的基本原理,并将隐密技术区分为约束式隐密技术和启发式隐密技术。其中,“约束式”是指以明确的隐密安全性理论为指导的隐密技术设计思想;“启发式”是指源于对隐密安全性经验性认识的设计思想。在此基础上,研究了统计安全性约束下的隐密模型,探讨了几个启发式隐密的关键问题,并分别采用直方图预膨胀策略和粒子群优化策略,提出了基于预处理的启发式隐密方法和基于统计安全性约束的隐密方法。在隐密分析模型及方法的设计方面,首先分析了现有隐密分析技术的设计思想,从方法论的视角将其分类为“概率演绎”隐密分析和“统计归纳”隐密分析(简称统计隐密分析)。进而在统计隐密分析范畴下给出了定性与定量隐密分析问题的数学描述,并应用有限样本集下的统计学习理论提出了相应的定性与定量隐密分析方法,包括基于载体分布边界的隐密判决方法、基于多类分类的数据来源判别方法、基于多元线性回归的隐密分析方法和基于支持向量回归的隐密分析方法。同时从统计特征整体性、集成学习方法和训练样本规模三个角度研究了进一步提高隐密分析性能的途径。更多还原

【Abstract】 For thousands of years, it has always been a significant research problem to seek secure ways for communicating sensitive information. With the rapid development of networks and multimedia technology, the ancient information hiding method has been rediscovered and received considerable attention, for concealing the very presence of secret communication in digital world is much easier. Consequently, the "hide and seek" problem under this secure communication paradigm, which is also called the "countermeasure between steganography and steganalysis", is becoming a hot topic for academic research in recent years. Aiming at the various problems in this field, a systematic study on steganographic countermeature has been made on the aspects of "security theory", "steganographic method" and "steganalytic method" respectively. The main contribution of this dissertation is as follows:As to the study of security theory, an analysis on the limitation of current theories and models on steganographic security is presented firstly, and then the difference between "stego data security" and "steganographic system (algorithm) securiy" is clearified. Under the perspective of comprehensive information theory, the connotation of steganographic security is rediscussed and a conclusion is drawn that steganography originates from the limited cognitive ability of subject to object in the sense of comprehensive information. Furthermore, a novel framework of steganographic security is constructed based on the concept of "cover distribution boundary", and some essential concepts are explicitly defined, including stego data security, steganographic system securiy and secure steganographic capacity of cover data. Finally, according to whether the steganographic algorithm is released or not, two security measurement models under this framework are presented, respectively.As to the steganographic model and method, the design principles of current typical steganographic technologies are analyzed at first. Moreover, heuristic steganography and constrained steganography are differentiated on the basis of different design ideas, with the former constrained by security measurement models and the latter originated from experiential understanding of security. In addition, a steganographic model constrained by statistical security is put forward and some key questions about heuristic steganography are discussed. Finally, a novel heuristic steganographic method and a novel constrained steganographic method are proposed, utilizing histogram preprocessing strategy and binary particle swarm optimization strategy respectively. As to the steganalytic model and method, different design ideas are summarized from the methodological point of view, and steganalytic technologies are reclassified as probabilistic deduction steganalysis and statistical induction steganalysis (statistical steganalysis for short). Furthermore, mathematical description of qualitative steganalysis and quantitative steganalysis are presented respectively under the category of statistical steganalysis. On the basis of these work, a series of novel steganalytic methods are proposed applying statistical learning theory in finite sample space, including steganalytic discriminant method based on cover distribution boundary, source forensic method based on multi-classification, quantitative steganalytic method based on multivariate regression and quantitative steganalytic method based on support vector regression. Finally, three approaches are studied to further improve steganalytic performance from the angles of holistic statistical features, ensemble learning and the scale of training database.更多还原