| 摘要: |
| 在最近的研究中，利用深度学习技术的载体选择式隐写方法取得了一些新进展。然而，这些方法仅仅将深度神经网络作为建立消息码字与图像之间映射关系的一种工具，没有考虑如何通过改进网络模型及优化特征分布以提升隐写方案的整体性能，导致它们在抗攻击性、通信容量以及完备性等关键指标上的表现难以满足实际通信场景的要求。为此，本文首先从映射规则设计方面入手，提出了一种基于优化神经网络特征嵌入的载体选择式隐写方法Feadio，具有非常好的完备性；其次，探究了原始图像与受攻击图像在嵌入空间中的分布关系，为了提高Feadio的抗攻击性与通信容量，更好地拉近属于同一类的受攻击图像与原始图像的特征嵌入距离，通过在超球面空间对嵌入特征分布进行优化，获得了更好的表征效果；最后，通过在训练模型时引入ArcFace Loss指标，有效地缩小了相同类别特征的间距，增大了不同类别特征的间距，使得Feadio中模型学习到的特征分布更具判别性。实验结果表明，Feadio不仅能保证100%的完备性和最先进的通信容量，在面对绝大多数几何和噪声攻击时可达到100%的抗攻击性，而且对于真实通信环境下总共12000张受攻击图像，获得了正确提取11997组图像消息的优秀表现。除此之外，本文首次提出了评估无载体隐写方法抗攻击性的基准数据集OSNA-Face，从更客观、更真实的角度衡量了方法的抗攻击性。Feadio的源代码和OSNA-Face数据集均可以从作者网站（https://ndsiiecas.github.io/）公开获取，以验证本文方法的有效性和实验结果的真实性。 |
| 关键词: 信息隐藏 载体选择式隐写术 深度度量学习 图像特征分布 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2024.04.05 |
| 投稿时间：2022-01-28修订日期：2022-04-25 |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| A coverless image steganography method using deep learning with feature distribution optimization |
|
WU Bin, XUE Rui
|
| (State Key Laboratory of Information Security,Institute of Information Engineering,Chinese Academy of Sciences) |
| Abstract: |
| In recent studies, selection-based coverless steganography using deep learning has made some new progress. How-ever, these methods only take the deep neural network as a tool to establish the mapping relationship between mes-sage codewords and images, and do not consider how to improve the performance of steganography schemes by improving the model and optimizing the feature distribution. As a result, their performance in key indicators such as resist attack, communication capacity and completeness are difficult to meet the requirements of practical commu-nication. Therefore, starting with the design of mapping rules, this paper proposes a selection-based coverless ste-ganography method Feadio based on optimized neural network feature embedding, which is complete. Secondly, the distribution relationship between original images and attacked images in the embedding space is explored. In order to improve the resist attack and communication capacity of Feadio, and shorten the feature embedding distance between the attacked images and the original images that belong to the same class, the better representation effect is obtained by optimizing the embedded feature distribution in the hypersphere space. Finally, by introducing the ArcFace Loss into the model, the distance of the same class features is effectively reduced and the distance of dif-ferent class features is increased, making the feature distribution learned by the model in Feadio more discrimina-tive. The experimental results show that Feadio can not only ensure 100% completeness and the most advanced communication capacity, but also achieve 100% resist attack in the face of most geometric and noise attacks, and obtain the excellent performance of correctly extracting 11997 messages for 12000 attacked images in the real communication. In addition, this paper proposes a benchmark dataset OSNA-Face to evaluate the ability of resist attack of coverless steganography method for the first time, which measures the resist attack ability of the method from a more objective and realistic point of view. The source code of Feadio and OSNA-Face dataset can be ob-tained from the authors" website (https://ndsiiecas.github.io/) to verify the effectiveness of this paper and the authenticity of the experimental results. |
| Key words: information hiding selection-based coverless steganography deep metric learning image feature distribution |
