基于生成式对抗网络的抗统计分析的高安全信息隐藏研究

Information hiding has been highly valued by governments and researchers in countries all over the world. However, information hiding currently faces two major challenges from statistical steganalysis and artificial intelligence aspects. To tackle these two challenges, this project, inspired by the ancient painter Cao Buxing's "dripping ink into a fly" anecdote, proposes that we can substantially change the cover image content at the regions in the cover image that are normally thought to be unsuitable for hiding information to hide information using the zero-sum game characteristics from Generative Adversarial Networks (GAN). Not only can this improve the information hiding capacity, but also can be resistant to statistical steganalysis. The project team will conduct in-depth research in the design theory, principles and methods of Information Hiding Generative Adversarial Networks (IH-GAN), in the GAN based information hiding resistant to steganalysis, in the multi-granular and multi-objective statistical steganalysis discriminative networks, in the highly secure information hiding deep learning models and their comparisons, in the security protection protocols of deep learning models, and so on. Simultaneously the team will develop verification platforms for information hiding based on GAN to validate the effectiveness of formed theories, methods and key technologies. During the execution of this project, some innovative theoretical results and key technological breakthroughs are achieved in terms of the automatic generation of information hiding target covers based on GAN, the extraction of multi-granular statistical features suitable for steganalysis, the parameter optimization of multi-granular and multi-objective IH-GAN etc. This research will have extremely important theoretical significance and application value for advancing our country's development in the field of information hiding resistant to statistical steganalysis.

信息隐藏深受各国政府与研究人员的重视，但它目前面临来自统计分析和人工智能的两大挑战。为了应对这两大挑战，本项目受古代画家曹不兴“落墨为蝇”思想的启发，提出利用生成式对抗网络（GAN）的博弈特点，在图像载体中通常认为不适合隐藏信息的区域，大幅度改变图像内容来隐藏信息；该方法不仅能提高隐藏容量，还能抵抗统计分析。项目组将对信息隐藏生成式对抗网络（IH-GAN）的设计理论，基于GAN的抗统计分析的信息隐藏，多粒度多目标统计分析判别网络，高安全信息隐藏深度学习模型及其比较等进行深入研究，并开发IH-GAN信息隐藏验证平台，对形成的理论、方法和关键技术进行有效性验证。通过项目的研究，在基于GAN的信息隐藏目标载体自动生成，适合隐写的多粒度统计特征提取，多粒度多目标IH-GAN参数优化等方面取得原创性理论成果和关键技术突破。该研究对我国在抗统计分析的信息隐藏领域发展将具有极其重要的理论意义与应用价值。

信息隐藏深受各国政府与研究人员的重视，但是面临统计分析和人工智能两大挑战。为了应对这两大挑战，本项目受古代画家曹不兴“落墨为蝇”思想的启发，提出利用生成式对抗网络（GAN）的博弈特点，在图像载体中通常认为不适合隐藏信息的区域，大幅度改变图像内容来隐藏信息；该方法不仅提高了隐藏容量，还能抵抗统计分析。项目组对信息隐藏生成式对抗网络（IH-GAN）的设计理论、基于GAN的抗统计分析的信息隐藏、多粒度多目标统计分析判别网络、高安全信息隐藏深度学习模型及其比较等进行了深入研究，并开发一款IH-GAN信息隐藏验证平台，对提出的理论、方法和关键技术进行有效性验证。通过项目的研究，在基于GAN的信息隐藏目标载体自动生成、适合隐写的多粒度统计特征提取、多粒度多目标IH-GAN参数优化等方面取得原创性理论成果和关键技术突破。该研究对我国在抗统计分析的信息隐藏领域发展将具有极其重要的理论意义与应用价值。在国内外重要期刊上发表论文81篇，其中IEEE TDSC等信息安全国际顶级期刊12余篇，SCI论文 80余篇。已申请国家发明专利17项，毕业研究生16名。项目相关成果获2019年度江苏省科学技术一等奖（基础类），项目负责人2019、2020年连续入选科睿唯安高被引学者。