【作者】 李阳；

【导师】 于丽芳；

【作者基本信息】 北京印刷学院 ， 电子信息， 2025， 硕士

【摘要】 随着信息技术的快速发展,信息安全已成为数字时代的一个关键课题。隐写术作为信息隐藏的重要手段,在保护数据安全方面发挥着积极作用,但同时也被不法分子用于恶意目的,对社会安全构成严重威胁。为应对这一挑战,隐写分析技术迅速发展,特别是深度学习方法的引入显著提升了检测性能,并克服了传统人工特征提取的局限性。然而,现有隐写分析方法主要面向实验室环境,当应用于实际场景时,由于训练数据(源域)与真实数据(目标域)之间的分布差异,模型性能往往出现显著退化,这个问题被称为载体源失配(Cover Source Mismatch,CSM)。针对这一关键问题,本文提出两种基于深度学习的隐写分析模型:(1)提出了一种基于可判别性感知的中间域的隐写分析网络GDNet。它包括可判别性混合区域生成(GDMR)和可判别性感知的局部图像混合(DLIM)两个模块,通过在区域级和像素级混合源图像和目标图像来构造有判别性的中间域,从而将在源域上训练的隐写分析器引导到目标域。一方面,GDMR通过设计一个与训练轮次相关的区域级混合比来控制混合区域的大小,并根据该混合比选择目标图像中与隐写信号强相关的区域参与中间域的生成,同时抑制与隐写信号弱相关的其他区域。另一方面,DLIM设置了像素级混合比调整机制,在区域级混合比逐步增加时,减小与隐写信号弱相关区域对中间域判别性的负面影响,进而增强中间域的多样性。理论分析和实验结果表明,GDNet在各种CSM场景下的预测性能均优于现有方法。(2)提出了一种基于生成式中间域的跨域相关性的失配隐写分析网络NCSNet。NCSNet构建了生成式中间域、通道维度跨域信息融合和空间维度跨域信息融合三个模块用于帮助隐写分析器从源域迁移到目标域。首先是一个分为三阶段生成式中间域,具体而言,首先通过使用噪声添加模块将噪声添加到目标样本以生成遵循源域分布的中间样本,随后对这些中间样本执行数据嵌入以生成隐写中间样本,来构建标记的中间域。在领域适应阶段提出两个模块集中于加强源域与目标域之间通道相关和空间相关的融合,首先是一个在通道维度上进行跨域信息融合的模块,通过融合两个域在通道上的信息,使得输出特征中能够同时包含源域和目标域的信息。其次考虑到预训练模型的先验知识,通过源域特征来引导目标域特征在空间位置上的学习,以此来提升目标域特征的可判别性。与目前的先进的隐写分析网络进行实验对比,证明了NCSNet在各种CSM场景下的有效性和优越性。更多还原

【Abstract】 With the rapid development of information technology,information security has become a key issue in the digital age.As an important means of information hiding,steganography plays a positive role in protecting data security,but it is also used by criminals for malicious purposes,posing a serious threat to social security.To address this challenge,steganalysis technology has rapidly developed,especially with the introduction of deep learning methods that significantly improve detection performance and overcome the limitations of traditional manual feature extraction.However,existing steganalysis methods are mainly aimed at laboratory environments.When applied to practical scenarios,the performance of the model often deteriorates significantly due to the distribution difference between the training data(source domain)and the real data(target domain).This problem is called Cover Source Mismatch(CSM).This article proposes two steganalysis models based on deep learning to address this critical issue:(1)A steganalysis network GDNet based on discriminative perception in the intermediate domain is proposed.It includes two modules:discriminative mixed region generation(GDMR)and discriminative aware local image mixing(DLIM),which construct discriminative intermediate domains by mixing source and target images at the region and pixel levels,thereby guiding the steganalysis trained on the source domain to the target domain.On the one hand,GDMR controls the size of the mixing region by designing a region level mixing ratio that is related to the epoch,and selects regions in the target image that are strongly correlated with the steganographic signal based on this mixing ratio to participate in the generation of the intermediate domain,while suppressing other regions that are weakly correlated with the steganographic signal.On the other hand,DLIM has set up a pixel level mixing ratio adjustment mechanism,which reduces the negative impact of weakly correlated regions with steganographic signals on the discriminative ability of the intermediate domain as the region level mixing ratio gradually increases,thereby enhancing the diversity of the intermediate domain.Theoretical analysis and experimental results indicate that GDNet outperforms existing methods in predicting various CSM scenarios.(2)A mismatch steganalysis network NCSNet based on generative intermediate domain cross domain correlation is proposed.NCSNet has built three modules:generative intermediate domain,channel dimension cross domain information fusion,and spatial dimension cross domain information fusion to assist steganalysis in migrating from the source domain to the target domain.Firstly,there is a three-stage generative intermediate domain.Specifically,the noise is added to the target sample using a noise addition module to generate intermediate samples that follow the distribution of the source domain.Then,data embedding is performed on these intermediate samples to generate steganographic intermediate samples,constructing the labeled intermediate domain.In the domain adaptation stage,two modules are proposed to focus on strengthening the fusion of channel and spatial correlations between the source domain and the target domain.Firstly,a module is used for cross domain information fusion in the channel dimension.By fusing the information of the two domains on the channel,the output features can simultaneously contain information from both the source and target domains.Secondly,considering the prior knowledge of the pre trained model,the source domain features are used to guide the learning of the target domain features in spatial position,thereby improving the discriminability of the target domain features.Experimental comparisons with advanced steganalysis networks have demonstrated the effectiveness and superiority of NCSNet in various CSM scenarios.更多还原