【作者】 李磊；

【导师】 段新涛；

【作者基本信息】 河南师范大学 ， 计算机科学与技术， 2022， 硕士

【摘要】 信息隐藏不同于密码学,其优势在于传递秘密信息的同时还可以掩盖通信秘密过程。大容量图像隐写术作为信息隐藏领域的重要分支,也随着计算机技术的进步成为了研究热点。现有的大容量图像隐写术大多是通过改进网络结构或引入新的模块来改进隐藏和提取效果,但基于现有隐写框架训练得到的模型生成的图像质量仍然有待提升,而且接收方很难判断载密图像是否来自于真正的发送方。针对这些问题,本文通过引入秘钥图像,利用神经网络“过拟合”的特性并结合差分图像叠加方法提出了一种新的图像隐藏框架。基于该框架训练得到的模型可生成质量更高的载密图像和提取秘密图像,且接收方可以轻松地判别载密图像是否来自于真正的发送方。其次,对前文中提出的框架进行优化并通过改进神经网络模型,进一步提升了隐写方案的容量和隐藏效果。本文研究内容可概括为如下两点:(1)针对现有基于深度卷积神经网络(Deep Convolutional Neural Network,DCNN)的图像隐藏框架训练得到的模型,所生成的载密图像和提取秘密图像的质量仍然有待提高的问题,本文提出了一种名为差分图像叠加(Difference Image Superposition Deep Hiding,DISDH)的基于DCNN的图像隐藏框架。首先,通过引入秘钥图像用于“过拟合”训练和差分图像获取,这会使得隐藏模型生成高质量的载密秘钥,但同时会降低载体的可选择性;其次,为了增加这种方案的载体可选择性,发送者使用载密秘钥减去秘钥图像得到差分图像并将该差分图像与一幅新的载体图像叠加得到载密图像。将载密图像通过公共信道传输至接收方;最后,接收者使用接收到的载密图像和原载体图像获取新的差分图像,并将其与秘钥图像叠加后作为提取网络的输入用于获取提取秘密图像。实验结果表明,在训练网络结构相同的神经网络模型时,相较于现有主流的基于DCNN的隐藏框架,基于DISDH训练得到的隐藏模型可以生成与原载体图像相似度更高的载密图像,提取模型可以提取出与原秘密图像相似度更高的提取秘密图像,而且接收者可轻易地判别出载密图像是否来自于真正的发送者。(2)为进一步提升隐藏容量,本文对DISDH进行了改进,将两幅与载体图像尺寸一致的秘密图像嵌入到一幅载体图像中,但是单纯地增加隐藏容量会降低隐藏模型的隐藏效果和提取模型的提取效果。为了解决增大容量给神经网络模型带来的负面影响,本文通过筛选秘钥图像的方式对DISDH进行了改进,同时又对隐藏网络模型和提取网络模型进行了优化。实验结果表明,基于改进的DISDH框架训练优化过的神经网络模型,在增大隐藏容量的同时,还可以保证模型生成高质量的载密图像和提取秘密图像。更多还原

【Abstract】 Information hiding is different to cryptography,the advantage of information hiding that it can hide the process of communication while transmitting information with secret.As an important branch of the field of information hiding,high-capacity image steganography has also become a research focus with the advancement of computer technology.The existing high-capacity image steganography mostly improves the hiding and reveal effects by improving the network structure or introducing new modules,but the quality of images generated by the models trained based on the existing steganography frameworks still needs to be improved,and for the receiver,it is very hard to distinguish whether the steganographic image is from the real sender.Aiming to solve these problems,a novel image hiding framework by introducing the key image,using the characteristic named “overfitting” of neural network and combining the method of difference image superposition is proposed in this article.The models trained based on the proposed framework can generate higher quality steganographic images and revealed secret images,and the receiver can easily determine whether the steganographic images are from the real sender.Secondly,the framework proposed in the front of this article is optimized and the hiding network and reveal network are improved as well,which aims to make the capacity higher and improve hiding effect of the steganography.The research content of this article can be summarized as the following two points:(1)Aiming to solve the problem that the quality of steganographic images and revealed secret images generated by the models trained based on the existing deep convolutional neural network(DCNN)based image hiding framework still needs to be improved.A novel DCNN-based image hiding framework called Difference Image Superposition Deep Hiding(DISDH)is proposed in this article.First,by introducing key image for “overfitting” training and difference image acquisition,“overfitting” training will enable the hiding model to generate high-quality key image with secret,but it will reduce the selectivity of cover images at the same time;secondly,in order improve to the selectivity of cover images that the receiver uses the difference image between the key image with secret and the original key image is superimposed on a new cover image to obtain the steganographic image.And it is transmitted to the receiver through the public channel;finally,the receiver uses the received steganographic image and the original cover image to obtain a new difference image,and superimposes it with the key image as the input of the reveal network to obtain the revealed secret image.The experimental results show that when training neural network models of hiding and reveal with same structures,compared with the existing DCNN-based hiding frameworks,the hiding models trained based on DISDH can generate steganographic images that are more similar to the original cover images,and the reveal model can obtain the revealed secret images which are more similar to the original secret images,and the receiver can determine whether the steganographic image is from the real sender easily.(2)In order to improve the capacity of hiding further,by improving DISDH and embeds two secret images with the same size as the cover image into a cover image in this article,but simply increasing the capacity of hiding will reduce the hiding effect of the hiding model and the reveal effect of the reveal model.In order to reduce the negative impact of increasing the capacity on the neural network models,DISDH is improved by screening key image,and the hiding network model and reveal network model are optimized at the same time in this article.The experimental results show that the optimized network models based on the improved DISDH framework can increase the capacity of hiding while ensuring that the models can generate high-quality steganographic images and revealed secret images.更多还原