基于CNN-Transformer的图像隐写分析  

Image steganalysis based on CNN-Transformer

作　　者：王嘉昊 阎红灿 谷建涛 WANG Jiahao;YAN Hongcan;and GU Jiantao(College of Science,North China University of Science and Technology,Tangshan 063210,Hebei Province,P.R.China;Key Laboratory of Data Science and Application of Hebei Province,Tangshan 063210,Hebei Province,P.R.China)

机构地区：[1]华北理工大学理学院,河北唐山063210 [2]河北省数据科学与应用重点实验室,河北唐山063210

出　　处：《深圳大学学报(理工版)》2025年第2期233-241,共9页Journal of Shenzhen University(Science and Engineering)

基　　金：河北省社会科学基金资助项目(HB17GL071);华北理工大学科研重点资助项目(ZD-YG-202316);河北省高等教育教学改革研究与实践资助项目(2023GJJG226)。

摘　　要：现有的卷积神经网络(convolutional neural network,CNN)隐写分析模型主要关注隐写图像的局部特征,尽管CNN通过堆叠更深的卷积层来扩展感受野,但在全局特征提取方面仍较为薄弱,而对于大尺寸图像,关注全局特征有助于提高隐写分析的效果.提出一种基于CNN-Transformer的图像隐写分析网络模型(CNN-Transformer image steganalysis network,CTS-Net),能够有效捕捉隐写信号局部特征与全局特征的依赖关系.在预处理阶段,采用多尺度残差提取与信息融合,以提高信噪比;在特征提取阶段,结合CNN和Transformer进行局部与全局特征提取,提高了对大尺寸隐写图像的检测精度;最后采用全连接层进行分类.在公开数据集BOSSbase1.01上采用WOW、HILL和S-UNIWARD隐写算法,使用不同嵌入率进行检测,结果表明,在低嵌入率(0.1 bpp)下,CTS-Net模型的检测准确率最优;在公开数据集ALASKA#2上采用WOW隐写算法,对16种图像尺寸的数据集进行隐写分析表明,CTS-Net模型在固定大小数据集与异构数据集上能有效利用隐写信号的全局特征,检测准确率高于SRNet、SiaStegNet、CvT Net和CVTS模型.Current convolutional neural network(CNN)steganalysis models primarily focus on the local features of steganographic images.Although CNNs expand their receptive field by stacking deeper convolutional layers,their ability to extract global features remains limited.For large images,focusing on global features can significantly improve steganalysis performance.We propose a hybrid model named CTS-Net(CNN-Transformer image steganography network)for image steganalysis.This model effectively captures both local and global features dependencies of the steganographic signals.In the preprocessing stage,multi-scale residual extraction and information fusion are applied to improve the signal to noise ratio.In the feature extraction stage,CNN and Transformer are combined to extract both local and global features,enhancing detection accuracy for large size steganographic images.Finally,a fully connected layer is used for classification.Experiments on the public dataset BOSSbase1.01,using the WOW,HILL,and S-UNIWARD steganography algorithms at different embedding rates for detection,show that at a low embedding rate(0.1 bpp),the CTS-Net model achieves the best detection accuracy.On the public dataset ALASKA#2,the WOW steganography algorithm is used for steganalysis across 16 image sets of different sizes.The results demonstrate that the CTS-Net model effectively leverages global features of the steganographic signals on both fixed-size and heterogeneous datasets,achieving superior detection accuracy compared to SRNet,SiaStegNet,CvT Net,and CVTS.

关 键 词：人工智能 卷积神经网络 图像隐写分析 多尺度残差 全局特征 大尺寸图像 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]