面向有损处理的图像鲁棒隐写关键问题研究

In order to save storage space, save traffic and protect intellectual property, instant messaging software, social networking sites and multimedia sharing sites often make some lossy processing to the transmitted images, such as lossy compressing, scaling, adding tags or watermark. This makes the information embedded by the traditional steganography can not be extracted correctly, and results in that people can not use the cover of massive image data in the Internet to implement covert communication. Therefore, the research on robust image steganography for lossy processing is urgently needed. This project aims to study the following problems in image robust steganography at present: 1) for a lot of lossy processing, there is not large capacity robust information embedding domain, 2) the existing distortion function can not accurately describe the performance of resistant to the existing steganalysis algorithms and 3) for a lot of lossy processing, there is not steganographic code which can correct error and minimize distortion at the same time. And the research will be focus on the following contents: constructing robust embedding domain based on image content feature, constructing robust embedding domain based on deep learning, designing image element changing distortion function based on parameter optimization, designing steganography changing distortion function based on modification mode selection, constructing decoding error model of STCs (Syndrome Trellis Codes) under lossy processing and designing steganography code which can minimize distortion and strongly correct error. Above researches are expected to improve the robustness and detection resistant performance, promote the practical application of steganography technology.

为了节约存储空间、节省流量、保护知识产权等目的，即时通信软件、社交网站和多媒体分享网站等往往会对传输的图像进行压缩、缩放、添加标签或者水印等有损处理。这使得传统隐写嵌入的信息无法正确提取，导致难以借助这些网站中大量图像数据的掩护实施隐蔽通信。因此，亟需开展面向有损处理的图像鲁棒隐写研究。本项目拟针对当前该方面研究中存在的下列问题开展研究：1）对很多有损处理缺乏大容量鲁棒嵌入域、2）现有失真函数无法准确刻画对已有隐写检测算法的抵抗能力和3）对很多有损处理缺乏兼顾纠错和失真最小化的隐写编码，重点研究基于图像内容特征的鲁棒嵌入域构建、基于深度学习的鲁棒嵌入域构建、基于参数优选的图像元素更改失真函数设计、基于修改模式优选的隐写更改失真函数设计、有损处理下STCs（网格编码）译码错误模型构建和强纠错最小失真隐写编码设计。上述研究有望提高隐写的鲁棒性和抗检测能力，促进隐写技术实用化。

互联网上各种平台、网站和工具中每天都传输着海量的图像数据。这些数据为实施隐蔽通信提供了丰富的载体资源。然而，为了节约成本、保护知识产权等，很多时候上传的图像都会遭受有损处理，导致现有很多高安全性的图像隐写算法嵌入的秘密信息无法被正确提取。为了能够更好地利用这些可能经过有损处理的图像作为掩护，亟待设计出对有损处理操作鲁棒的图像隐写算法。针对该问题，本项目从鲁棒的嵌入域构建、隐写更改失真函数设计、隐写编码等三个环节入手，开展面向有损处理的图像鲁棒隐写关键问题研究，提出了3种基于变换域SVD的图像鲁棒隐写算法和2种基于深度神经网络的图像鲁棒信息隐藏算法；在此基础上进一步拓展，提出了2种抗H.264压缩的视频鲁棒隐写算法和1种鲁棒网络流水印算法；提出了2种分别适用于雾天图像和多载体图像的高安全性隐写算法，以及基于STC级联CRC码的隐写编码方法和基于信息分布差异的带矩阵编码F5隐写密钥还原方法；最后，在反向隐写检测方面，提出了2种自适应图像隐写检测算法、2种高动态图像隐写分析算法；在隐写信息定位方面，给出了基于定量隐写分析的隐写信息定位一般方法、基于同频子图的JPEG图像隐写信息定位框架，以及3种适用于空域图像和JPEG图像的隐写信息定位算法。.上述结果已经形成论文，并在IEEE TNSM、Signal Processing和“计算机学报”等国内外重要期刊和学术会议上发表相关论文33篇，其中SCI期刊26篇；申请并获国家发明专利授权3项。培养博士生1名，硕士生6名。研究结果已经转化成日常教学内容，研制的系统也已在相关部门应用，并取得了一定成效。