JPEG图像失配隐密分析研究
发布时间：2019-01-27 07:34
【摘要】：随着网络通信技术的发展,以隐蔽通信为目的的隐密术受到了社会的广泛关注。隐密术是指将秘密信息嵌入到载体数据的冗余位置,利用公开信道以不被察觉的方式进行秘密通信的技术。虽然隐密术在隐密通信和知识产权保护等方面给社会提供了便利,但也被不法分子应用到隐蔽地传输消息等方面,给社会的安全带来了严重的威胁。因此,研究如何从公共信道的海量数据中识别出含有秘密信息的文件的隐密分析技术具有重要的现实意义。 传统隐密分析依赖于已经获得隐密者的载体样本和含密样本的假设。而在实际应用中,该假设通常不会满足,使得隐密分析会出现失配问题。目前,虽然已经有许多文献指出失配问题会导致传统隐密分析算法性能下降,但鲜有有效的算法能够克服隐密分析中的失配问题。本文从机器学习角度出发,研究了失配因素对于传统隐密分析特征的影响,针对不同的应用环境,提出了基于局部领域泛化的融合训练失配隐密分析以及基于迁移学习的失配隐密分析广义转移成分分析方法。本文的研究成果如下： (1)首先介绍了传统隐密分析框架,包括研究背景和意义、基本概念和研究现状,重点介绍了几种典型的隐密分析特征和常用的机器学习工具。其次,给出失配隐密分析框架,分别从载体图像生成过程和含密图像生成过程讨论了不同的失配因素对于传统隐密分析性能的影响。最后,总结了前人失配隐密分析的研究方法,分析了各个方法的应用环境和优缺点。 (2)通过总结前人基于融合训练失配隐密分析的策略,提出基于局部领域泛化的融合训练方法。该方法引入待测图像局部领域的概念,通过降低待测图像局部领域特征分布的方差,并且保持局部领域训练数据和标签的相关性,提取能够泛化局部领域的公有特征,利用该特征对待测图像进行隐密分析。将该融合训练方法与前人融合训练方法进行比较,在相同的实验环境下,失配隐密分析的判决错误率降低了2%-6%。 (3)针对基于融合训练失配隐密分析方法训练数据的多样性需求的局限性,引入迁移学习的思想,提出广义转移成分分析失配隐密分析方法。此方法可以根据不同的测试图像,自适应的调整单源的训练库数据的特征分布,使得其能应用于多种失配因素的隐密分析检测。通过与前人方法的比较,该方法能够在有限的单源的训练数据的情况下,使得其达到与融合训练同等级的失配隐密分析的性能。此外,该方法对于多种失配因素具有鲁棒性。
[Abstract]:With the development of network communication technology, covert communication has been paid more and more attention. Steganography refers to the technology of embedding secret information into the redundant position of carrier data and using open channels to communicate secretly in an undetected manner. Although secret technology provides convenience to society in secret communication and intellectual property protection, it is also used by lawless elements to transmit information in secret, which brings serious threat to the security of society. Therefore, it is of great practical significance to study how to identify the secret analysis technology of files containing secret information from the mass data of common channels. Traditional cryptographic analysis relies on the assumption that carrier samples and secret samples have been obtained. However, in practical application, the assumption is usually not satisfied, which leads to the mismatch problem. At present, although many literatures have pointed out that the mismatch problem will lead to the performance degradation of the traditional secret analysis algorithm, there are few effective algorithms to overcome the mismatch problem in the secret analysis. From the point of view of machine learning, this paper studies the influence of mismatch factors on the characteristics of traditional secret analysis, aiming at different application environments. In this paper, a fusion training mismatch secret analysis based on local domain generalization and a generalized transfer component analysis method based on transfer learning are proposed. The results of this paper are as follows: (1) this paper firstly introduces the traditional secret analysis framework, including the research background and significance, basic concepts and research status, focusing on the introduction of several typical features of secret analysis and commonly used machine learning tools. Secondly, the mismatch secret analysis framework is given, and the influence of different mismatch factors on the performance of traditional secret analysis is discussed from the process of image generation and the process of generating secret image respectively. Finally, the research methods of mismatch secret analysis are summarized, and the application environment, advantages and disadvantages of each method are analyzed. (2) by summing up the previous strategy of mismatch secret analysis based on fusion training, a fusion training method based on local domain generalization is proposed. This method introduces the concept of the local domain of the image under test. By reducing the variance of the local domain feature distribution of the image under test, and keeping the correlation between the local domain training data and the label, the public features of the local domain can be generalized. The feature is used for secret analysis of the measured image. The fusion training method is compared with the previous fusion training method. Under the same experimental environment, the error rate of mismatch secret analysis is reduced by 2%-6%. (3) aiming at the limitation of diversity requirement of training data based on fusion training mismatch secret analysis method, a generalized transfer component analysis mismatch secret analysis method is proposed by introducing the idea of transfer learning. This method can adaptively adjust the feature distribution of single source training database data according to different test images, so that it can be applied to the hidden analysis and detection of many mismatch factors. Compared with previous methods, this method can achieve the performance of mismatch secret analysis with the same level of fusion training under the condition of limited single source training data. In addition, the method is robust to various mismatch factors.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP309.7
本文编号：2416051
[Abstract]:With the development of network communication technology, covert communication has been paid more and more attention. Steganography refers to the technology of embedding secret information into the redundant position of carrier data and using open channels to communicate secretly in an undetected manner. Although secret technology provides convenience to society in secret communication and intellectual property protection, it is also used by lawless elements to transmit information in secret, which brings serious threat to the security of society. Therefore, it is of great practical significance to study how to identify the secret analysis technology of files containing secret information from the mass data of common channels. Traditional cryptographic analysis relies on the assumption that carrier samples and secret samples have been obtained. However, in practical application, the assumption is usually not satisfied, which leads to the mismatch problem. At present, although many literatures have pointed out that the mismatch problem will lead to the performance degradation of the traditional secret analysis algorithm, there are few effective algorithms to overcome the mismatch problem in the secret analysis. From the point of view of machine learning, this paper studies the influence of mismatch factors on the characteristics of traditional secret analysis, aiming at different application environments. In this paper, a fusion training mismatch secret analysis based on local domain generalization and a generalized transfer component analysis method based on transfer learning are proposed. The results of this paper are as follows: (1) this paper firstly introduces the traditional secret analysis framework, including the research background and significance, basic concepts and research status, focusing on the introduction of several typical features of secret analysis and commonly used machine learning tools. Secondly, the mismatch secret analysis framework is given, and the influence of different mismatch factors on the performance of traditional secret analysis is discussed from the process of image generation and the process of generating secret image respectively. Finally, the research methods of mismatch secret analysis are summarized, and the application environment, advantages and disadvantages of each method are analyzed. (2) by summing up the previous strategy of mismatch secret analysis based on fusion training, a fusion training method based on local domain generalization is proposed. This method introduces the concept of the local domain of the image under test. By reducing the variance of the local domain feature distribution of the image under test, and keeping the correlation between the local domain training data and the label, the public features of the local domain can be generalized. The feature is used for secret analysis of the measured image. The fusion training method is compared with the previous fusion training method. Under the same experimental environment, the error rate of mismatch secret analysis is reduced by 2%-6%. (3) aiming at the limitation of diversity requirement of training data based on fusion training mismatch secret analysis method, a generalized transfer component analysis mismatch secret analysis method is proposed by introducing the idea of transfer learning. This method can adaptively adjust the feature distribution of single source training database data according to different test images, so that it can be applied to the hidden analysis and detection of many mismatch factors. Compared with previous methods, this method can achieve the performance of mismatch secret analysis with the same level of fusion training under the condition of limited single source training data. In addition, the method is robust to various mismatch factors.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP309.7
【参考文献】
相关期刊论文 前1条
1 黄炜;赵险峰;盛任农;;基于KFD指标聚类的高隐蔽性JPEG隐写分析[J];计算机学报;2012年09期
相关博士学位论文 前1条
1 郭艳卿;隐密对抗的理论及方法研究[D];大连理工大学;2009年
，本文编号：2416051
本文链接：https://www.wllwen.com/falvlunwen/zhishichanquanfa/2416051.html
