| 引用本文： |
-
赵枫,赵险峰,易小伟,何晓磊,肖俊超.基于信道选择和深度特征融合的大尺寸图像隐写分析方法研究[J].信息安全学报,已采用 [点击复制]
- zhaofeng,zhaoxianfeng,yixiaowei,hexiaolei,Xiaojunchao.Steganalysis of large-size image based on channel selection and deep feature fusion[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 近年来，自适应图像隐写技术的发展给图像隐写分析技术带来了很大的挑战，而深度学习的相关技术在隐写分析领域的应用使隐写分析技术取得了突破性进展，目前各方面性能均已超过机器学习隐写分析技术。但是由于GPU受限，当前图像隐写分析网络可分析图像的尺寸仍局限在较小尺寸范围内，无法对较大尺寸的图像直接分析，随着多媒体技术的不断发展，大尺寸、高分辨率的图像已经常态化使用，绝大多数隐写分析网络已经不再适应当前的环境，针对该问题本文提出了一种基于信道选择和深度特征融合的大尺寸图像隐写分析方法，将较大尺寸图像分成网络可直接训练和检测的尺寸的图像块，并设计了一种计算特征复杂度的方法，然后基于对现有网络的改进使网络同时输出分块图像的隐写判别结果和提取深度特征的复杂度，再根据内容自适应算法的基本原理，利用选择信道知识自适应提供给融合判别器分块图像的权重，最终对所有分块图像的检测结果和自适应权重融合判决。为了保证隐写信息的不被破坏，本文对图像采用了重叠分块，本文方法在分析较大尺寸图像上取得了较高的检测准确率，当前环境下具有比较高的实用性，实验证明，在多种自适应隐写算法上本文方法在目前深度学习网络可直接分析的略大尺寸图像上取得了比直接分析方法更优的检测性能，并在当前网络无法直接分析的较大尺寸图像上取得了优于传统的融合信道选择知识的通用隐写分析方法的检测性能，检测准确率最高可提升10%。
关键词 隐写分析；深度学习；大尺寸图像；特征复杂度；信道选择；深度特征融合 |
| 关键词: 隐写分析 深度学习 大尺寸图像 特征复杂度 信道选择 深度特征融合 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2024.02.25 |
| 投稿时间：2022-09-06修订日期：2022-12-15 |
| 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） |
|
| Steganalysis of large-size image based on channel selection and deep feature fusion |
|
zhaofeng, zhaoxianfeng, yixiaowei, hexiaolei, Xiaojunchao
|
| (State Key Laboratory of Information Security, Institute of Information Engineering, Chinese Academy of Sciences) |
| Abstract: |
| In recent years, the development of adaptive image steganography technology has brought great challenges to image steganalysis technology, and the application of deep learning in the field of steganalysis has made breakthrough pro-gress in steganalysis technology. All aspects of performance have surpassed machine learning steganalysis technology. However, due to the limitation of GPU, the size of images that can be analyzed by the current image steganalysis net-work is still limited to a small size range, and it is impossible to directly analyze images of larger sizes. With the devel-opment of multimedia technology, large-size, high-resolution images have been used normally, and most steganalysis networks are no longer suitable for the current environment. To solve this problem, this paper proposes a large-size im-age steganalysis method based on channel selection and deep feature fusion. The image is divided into image blocks of the size that the network can directly train and detect, and a method to calculate the feature complexity is designed. Then based on the improvement of the existing network, the network can output the steganographic discrimination results of the block image and extract the depth features at the same time. Then, according to the basic principle of the content adaptive algorithm, the weight of the block image provided to the fusion discriminator is adaptively provided by the knowledge of the selection channel. Finally, use the detection results of all the block images and the adaptive weight to make a fusion decision. This method can achieve high detection accuracy on larger sizes images , which is relatively high in the current environment. Experiments have shown that the method has achieved better detection performance than the direct analysis method on a variety of adaptive steganography algorithms on slightly larger images that can be directly analyzed by the current network. On the analyzed large-size images, the detection performance is better than that of the traditional general steganalysis method fused with channel selection knowledge, and the detection accuracy can be improved by up to 10%. |
| Key words: steganalysis deep learning feature fusion larger size image deep feature fusion Complexity of features |
