基于张量分解框架的深度学习信息隐藏对抗研究

As an important security technology for protecting contents in communication, the research on information hiding warfare has great significance both in academics and applications. Deep learning, as a booming technology, is an attractive solution to challenges and issues in the process that the research on information hiding warfare moves from lab environment towards real world. However, lacking of strict theoretical foundation and easily falling into local minimum, these two shortcomings severely restrict the application of deep learning in information-hiding warfare. This project introduces tensor decompression framework into the research of deep-learning based information-hiding warfare. We establish tensor-decompression based latent-variable cover distribution model. Based on it, we construct deep-learning steganalysis model and deep-learning minimal distortion steganography model, respectively. Our goal is to establish non-additive distortion measure which can utilize the feedback of deep-learning steganalytic detector, deduce the risk bounds of deep-learning steganalytic model in theory and propose a tensor-decompression-framework based solution to cover source mismatching. Via this project, we try to establish deep-learning based information-hiding joint confrontation and optimization theory and implementation mechanism which is applicable to small dataset, possesses transferability, and avoids falling into local optimal solutions in the sample space. It is expected the research results will provide the theoretical and technical support for the applications of steganography and steganalysis.

作为保障通信安全的一种重要手段，信息隐藏与对抗的研究具有重大的学术意义和应用价值。针对目前信息隐藏对抗研究从实验室环境迈向实际应用过程中存在的问题和挑战，深度学习提供了一种很好的方法。但是，缺乏严格的理论支撑以及在实践中容易陷入局部次优解，这两大缺陷极大地制约了深度学习在信息隐藏对抗中的发展。本项目把张量分解理论框架引入深度学习信息隐藏对抗研究中，建立基于张量分解框架的载体潜变量分布模型，并在此基础上分别构建深度学习隐写分析模型以及深度学习最小失真隐写框架；实现能利用隐写分析器的反馈进行对抗学习的非加性失真度量；对隐写分析模型的分类风险上界进行理论分析，并提出在张量分解框架下的载体源失配问题解决方案。通过研究，建立能适用于小数据集、具备可迁移性、避免陷入样本空间局部次优解的深度学习信息隐藏对抗优化理论和实现机制，为信息隐藏对抗的实际应用提供理论和技术支撑。

作为保障通信安全的一种重要手段，信息隐藏与对抗的研究具有重大的学术意义和应用价值。针对目前信息隐藏对抗研究从实验室环境迈向实际应用过程中存在的问题和挑战，本项目通过把隐写及隐写分析模型进行张量分析和张量分解，研究能适用于小数据集、具备可迁移性、避免陷入样本空间局部次优解的深度学习信息隐藏对抗优化理论和实现机制。通过项目资助，本项目提出了适用于空域、JPEG域、真彩色图像，行之有效的深度学习隐写分析模型；提出了基于强化学习的非加性隐写失真度量；并将信息隐藏对抗领域行之有效的技术应用于图像视频等多媒体取证领域。依托本项目，在国内外重要期刊和学术会议发表论文26篇，其中ACM/IEEE Transactions 以及中国计算机学会A 类期刊/A 类会议论文8篇。SCI论文17篇。申请发明专利11项。培养研究生11人，其中博士生5人，2人获得博士学位，硕士生6人，6人获得硕士学位。项目组2人晋升高级职称。项目负责人作为主要完成人（排名第四）的“信息隐藏理论与方法”项目荣获2019年中国计算机学会科学技术奖自然科学一等奖。本项目涉及的隐写及隐写分析技术有望应用于国防安全、公安鉴证取证、商业防伪等多个领域。