针对在公共数据平台分享隐写数据进行隐蔽通信的隐写分析方法研究

Sharing stego media data on public data sharing platforms such as SNS can be used for secret communication. Comparing to transmitting stego data in conventional point-to-point channel, this method arouses less suspicion to communication behavior and also confuses detector by huge number of false alarm on innocent data presented on public data sharing platforms, thus it is of higher security and practicability. This kind of stego communication differs from conventional method by steganographic embedding scheme and communication behavior. Specifically, it has to design robust steganographic embedding method to overcome JPEG recompression on uploaded images by public data sharing platform, and meanwhile maintain its security and payload capacity. Besides, it can adopt conspiracy schema via sharing stego data among multiple users to diversify stego data transmission and disguise secret communication behavior. This project aims at building a protection framework for finding and preventing illegal stego communication on public data sharing platforms by detecting both this kind of stego data and stego communication user behavior. In detection of robust stego data, we first analyze the common characteristics of current robust stego embedding methods and then associate them with steganalysis feature design for such stego data. For this purpose, we investigate several issues including multiple-object selecet-channel aware feature construction, high order tensorized feature representation and tensor decomposition for feature design, and novel statistical detection model incorporating channel recompression knowledge. Based on stego data detection and in order to improve detection accuracy, we detect steganographic users and their conspiring communication group by simultaneously utilizing stego data detection results and modeling data sharing interactions among users via graph convolutional networks. Distinguishing stego communication users and their conspiring group are described as node classification problem for graph data, and we further consider and model the time variant properties of stego data transmission flow in the graph to enhance its detecting ability. In this project, targeted detection on robust steganographic embedding and detection on steganographic users and conspiring communication group are closely related and integrated, which means the former lies a foundation for the later, and the later provide higher level detection with higher accuracy. In this project, research works of the two parts stem from characterizing the real application problems, thus they have both theoretical and applicable prospects.

利用社交网络等公共数据平台分享隐写图像进行隐蔽通信具有较高的安全性和实用性。与传统的端对端信道传输隐写数据相比，其通信行为更为隐蔽，隐写数据可被大量正常数据的虚警所混淆。该类隐写方法需要设计鲁棒隐写方法以克服平台对上传图像重压缩带来的影响并同时保证其安全性和容量，此外可能多人合谋分享进行隐写数据分散传递和行为掩护。本项目针对此类隐写数据和用户行为研究检测方法，提供防范措施。其中，隐写数据检测主要围绕该类鲁棒隐写方法共性特点，研究隐写嵌入对于图像重压缩特性的影响，以此构建多目标选择信道特征，并结合高阶张量特征构造和分解方法以及新型的隐写检测模型，提升检测性能。基于隐写数据检测信息，利用图数据描述用户间数据交互，构建图数据的卷积神经网络和图节点分析方法研究隐写用户及其合谋团体检测，同时结合隐写数据流转的时变关系提升检测精度。本项目针对该问题联合隐写数据和行为的特点建模研究，具有理论和应用价值。

利用社交网络等公共数据平台分享隐写图像进行隐蔽通信具有较高的安全性和实用性。与 传统的端对端信道传输隐写数据相比，其通信行为更为隐蔽，隐写数据可被大量正常数据的虚警所混淆。该类隐写方法需要设计鲁棒隐写方法以克服平台对上传图像重压缩带来的影响并同 时保证其安全性和容量，此外可能多人合谋分享进行隐写数据分散传递和行为掩护。本项目针对此类隐写数据和用户行为研究检测方法，提供防范措施。其中，隐写数据检测主要围绕该类鲁棒隐写方法共性特点，研究隐写嵌入对于图像重压缩特性的影响，以此构建新型的隐写检测模型，提升检测性能。基于隐写数据检测信息，利用图数据描述用户间数据交互，构建图数据的卷积神经网络和图节点分析方法研究隐写用户及其合谋团体检测。此外，本项目还研究了针对少量数据样本训练条件下的隐写分析方法，以及信道知识在鲁棒隐写编码和纠错译码中的应用。