| 引用本文： |
-
陈屹东,赵险峰,何晓磊,刘长军.基于全卷积网络的图像盲隐写分析与算法判别方法[J].信息安全学报,已采用 [点击复制]
- chenyidong,zhaoxianfeng,hexiaolei,liuchangjun.Image blind steganalysis and algorithm distinguishing based on fully convolutional network[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 隐写是一种不同于加密的信息隐藏技术。隐写技术利用载体的信息冗余隐藏秘密消息，使得传输行为不被发现，是对传统加密手段的补充。近年来，随着互联网上图像载体日益增长，以图像为载体的隐写方法在日常生活中被越来越多地使用。隐写分析是隐写的对抗面，目的是发现隐写行为的存在并破解隐写信息。为了防止隐写技术的滥用，隐写分析技术也需要在应用环境中得到新的发展。在应用环境中的隐写分析往往对隐写载体和算法缺乏先验知识，这被称为盲隐写分析。在应用环境中，隐写者可能使用自适应隐写算法嵌入信息，也可能使用非自适应隐写算法，这两者嵌入原理有所差异，对应的高效分析方法也完全不同。因此盲隐写分析中如何确定隐写方使用算法是一大难题，并具有实际意义。本文基于全卷积网络和自适应隐写算法的特点，提出一个对针对自适应与非自适应混合图像样本的盲隐写分析框架。该框架根据全卷积网络输出每个样本各像素点隐写概率，并根据判定的隐写点分布计算区域嵌入复杂度和嵌入纹理复杂度，结合两者对样本隐写算法联合判别。该框架只需在自适应隐写样本上训练，并可以接受任意尺寸的图像输入。本文在空域和JPEG域均选择了自适应与非自适应隐写算法，构建样本数据集，对所提框架的隐写分析和算法判别能力均进行了实验。在标准图像库BOSSBase上的实验验证了本文框架的有效性。 |
| 关键词: 全卷积网络 盲隐写分析 隐写算法 |
| DOI： |
| 投稿时间：2022-07-31修订日期：2022-09-24 |
| 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）,国家重点基础研究发展计划（973计划） |
|
| Image blind steganalysis and algorithm distinguishing based on fully convolutional network |
|
chenyidong, zhaoxianfeng, hexiaolei, liuchangjun
|
| (State Key Laboratory of Information Security, Institute of Information Engineering) |
| Abstract: |
| Steganography is an information hiding technique that is different from encryption. More specifically, steganogra-phy is a supplement to traditional encryption, which hides secret messages by using the information redundancy of the cover. In recent years, with the increasing number of image covers on the Internet, steganography based on im-ages has been used more and more in daily life. Steganalysis is the antagonistic aspect of steganography, which aims to discover the existence of steganography behavior and crack the steganography information. In order to prevent the abuse of steganography, steganalysis should also be developed in the application environment. Steganalysis of-ten lacks prior knowledge of steganography covers and algorithms in the application environment, which is called blind steganalysis. In the application environment, steganographers may use adaptive steganographic algorithm to embed information, or they may use non-adaptive steganographic algorithm. The embedding principles of these two algorithms are different, and the corresponding efficient steganalysis methods are completely different. Therefore, how to determine the steganographic algorithm in steganalysis is a difficult problem and has practical significance. Based on the characteristics of fully convolutional network and adaptive steganographic algorithm, a blind ste-ganalysis framework for adaptive and non-adaptive hybrid image samples is proposed in this paper. The framework outputs the steganography probability of each pixel of each sample according to the fully convolutional network, calculates the regional embedding complexity and the embedding texture complexity according to the distribution of the judged steganography points, and combines the two to make a joint judgment on the steganographic algorithm of the input sample. This framework only needs to be trained on samples generated by adaptive steganographic algo-rithm and can accept image input of any size. In this paper, adaptive and non-adaptive steganographic algorithms are selected in both spatial and JPEG domains to construct sample datasets, and the steganalysis and algorithm discrim-ination ability of the proposed framework are tested. Experiments on the standard image dataset BOSSBase verify the validity of the proposed framework. |
| Key words: fully convolutional network blind steganalysis steganographic algorithm |
