引用本文
  • 刘婷,任延珍,王丽娜.基于条件可逆网络的生成式图像隐写算法[J].信息安全学报,2023,8(4):17-30    [点击复制]
  • LIU Ting,REN Yanzhen,WANG Lina.Generative Image Steganography Via Conditional Invertible Neural Network[J].Journal of Cyber Security,2023,8(4):17-30   [点击复制]
【打印本页】 【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览 7972次   下载 5759 本文二维码信息
码上扫一扫!
基于条件可逆网络的生成式图像隐写算法
刘婷, 任延珍, 王丽娜
0
(武汉大学国家网络安全学院 武汉 中国 430072)
摘要:
近年来,随着生成模型的广泛使用,生成式隐写领域得到了快速发展。生成式隐写是在图像合成过程中隐藏信息的技术。它无需真实图像参与,只需秘密消息驱动生成模型即可合成载密图像。然而,现有方法无法控制生成的图像内容,因此不能保证隐蔽通信行为的安全性。针对上述问题,本文提出了基于条件可逆网络(Conditional Invertible Neural Network,cINN)的生成式图像隐写术steg-Cinn。在本文中,我们将信息隐藏建模为图像着色问题,并将秘密信息嵌入到灰度图像的颜色信息中。首先,我们使用映射模块将二进制秘密信息转换为服从标准正态分布的隐变量。而后,我们以灰度图像作为先验来指导着色过程,使用条件可逆网络来将隐变量映射为颜色信息。其中steg-Cinn生成的彩色图像匹配灰度图像的语义内容,从而保证了隐蔽通信的行为安全。对比实验结果表明,本文方法能够控制生成的图像内容并且使得合成颜色真实自然,在视觉隐蔽性方面表现良好。在统计安全性方面,本文方法的隐写分析检测正确率为56.28%,说明它能够抵御隐写分析检测。此外,本文方法在比特消息提取方面可以实现100%正确提取,这种情况下的隐藏容量是2.00 bpp。因此,与现有方法相比,本文方法在图像质量、统计安全性、比特提取正确率和隐藏容量方面取得了良好的综合性能表现。迄今为止,本文方法是在图像隐写术中首次使用cINN的工作。考虑到任何信息都可以转换为二进制形式,我们可以在图像中隐藏任意类型的数据,因此本文方法在现实世界里也具备实用价值。
关键词:  图像隐写|可逆网络|内容可控|行为安全
DOI:10.19363/J.cnki.cn10-1380/tn.2023.07.02
投稿时间:2021-12-18修订日期:2022-02-19
基金项目:本文受到国家自然科学基金项目支持(No. 62172306) 以及湖北省重点研发计划项目(No. 2021BAA034)支持。
Generative Image Steganography Via Conditional Invertible Neural Network
LIU Ting, REN Yanzhen, WANG Lina
(School of Cyber Science and Engineering, Wuhan University, Wuhan 430072, China)
Abstract:
In recent years, the field of generative steganography has developed rapidly with the widespread use of generative models. Generative steganography is a technology of generating stego images directly from secret messages without real images. However, existing works can't control the generated image content, thus can't guarantee behavioral security during covert communication. To address the above issue, this paper proposes steg-Cinn, a generative image steganography based on Conditional Invertible Neural Network(cINN). In this paper, we formulate the data hiding as an image colorization problem and the secret data is embedded into the color information for a gray-scale host image. First, the binary secret data is transformed into latent variable that follows standard normal distribution using a mapping module. Second, we use a conditional invertible neural network which uses gray-scale image as prior to guide the colorization process, where the latent variable is mapped into color information. The colored images generated by steg-Cinn can match semantic content of gray-scale images, thus ensuring behavioral security in covert communication. The comparative experimental results show that the proposed method is able to control the generated image content and generate realism colors, indicating good performance in terms of visual concealment. For statistical security, the proposed method can resist the detection of steganalysis successfully, where the detection rate calculated by steganalyzer is 56.28 %. In addition, the proposed method can achieve 100% bit extraction accuracy with the hiding capacity of 2.00 bits per pixel(bpp). Therefore, comparing with the existing methods, the proposed method achieves good comprehensive performance in terms of image quality, statistical security, bit extraction accuracy and hiding capacity. As far as we know, the proposed method is the first work to use cINN in image steganography. Since any information can be binarized, we can embed data with arbitrary types into images, thus bringing practical utility in the real-world.
Key words:  image steganography|invertible neural network|content controllability|behavioral security