摘要
近年来,深度学习在图像隐写分析任务中表现出了优越的性能。然而,此类方法在捕获图像中微弱的隐写噪声时,往往会因下采样过程中大量关键细节信息的丢失,导致在检测空域和JPEG隐写图像时难以同时实现高检测准确率。为此,本文基于孪生神经网络对图像进行分区域细粒度学习,同时利用交叉注意力机制进一步增强模型全局信息感知能力,提出一种跨通道交叉注意力增强的隐写分析方法(CES-Net)。首先,采用孪生神经网络作为主干网对图像进行分区域学习,以细致地感知空域和JPEG图像的像素信息和微弱的隐写噪声,同时,设计了多样化的高通滤波器和多层卷积作为网络预处理层来获取丰富且高质量的隐写噪声残差;接着,改进了特征提取部分,提出了跨通道交叉注意力网络,使模型提取到更多因隐写嵌入对图像像素相关性造成扰动的隐写特征,用于基于秘密噪声残差等弱信息的隐写图像分类任务;最后,融合子网络学习到的不同区域图像的分类特征,并输入全连接层组成的分类模块对载体和载密图像进行分类,提升检测效果。在隐写和隐写分析领域常用的图像数据集BOSSBase-1.01和BOWs2上进行了大量实验,结果表明,CES-Net方法与现有方法相比,对于空域和JPEG图像的多种主流隐写算法均能达到目前最优的检测准确率,其中,对多种空域隐写算法(WOW、S-UNIWARD和HILL)在不同嵌入比率下生成的载密图像,检测准确率最高分别提升1.27%~25.61%、2.1%~21.73%和1.69%~23.46%;对JPEG图像自适应隐写算法J-UNIWARD在不同嵌入比率下生成的载密图像,CES-Net方法对两种质量因子(QF=75和QF=85)的JPEG图像隐写检测准确率最高分别提升2.34%和2.06%。
In recent years,deep learning has demonstrated excellent performance in image steganalysis.The deep learning-based steganalysis method constructs an end-to-end network through a data-driven approach for classification tasks,thereby significantly reducing human intervention and achieving outstanding detection performance.However,these methods frequently encounter challenges in achieving high detection accuracy when capturing subtle steganography noise in images.This is primarily due to the substantial loss of critical details during the down-sampling process,which significantly hinders the detection of both spatialdomain and JPEG image steganography.For this reason,this paper presents a comprehensive analysis of the mechanism through which different modules of the steganalysis network effectively capture steganography noise residuals,and proposes CES-Net,a novel steganalysis method.CES-Net leverages Siamese neural networks for fine-grained regional learning of images and incorporates a cross-attention mechanism to further enhance the model's ability to perceive global information.Firstly,we employ a Siamese neural network as the backbone for regional learning on images,enabling precise perception of pixel information and subtle steganography noise in spatial-domain and JPEG-compressed images.This not only aids in uncovering hidden information within the image but also markedly enhances the model's sensitivity to specific patterns or features,thereby leading to more accurate and reliable results.In the pre-processing module,we design diverse high-pass filters and multi-layer convolutions to enhance the model's ability in extracting steganography noise residuals from different regions of the image.This provides abundant and high-quality steganography noise residuals for the subsequent network's learning process.Secondly,we propose incorporating a cross-channel cross-attention network into the feature extraction module.This enhances the model's ability to extract steganography features that have been disrupted by embedding,thereby enabling accurate classification of stego images based on subtle information such as secret noise residuals.In this module,CES-Net introduces a novel block that leverages the cross-attention mechanism.This block can effectively capture both intra-channel and inter-channel correlations in residual maps,thereby enhancing the representational capacity of steganography noise features and making the extracted features more discriminative.Finally,at the end of the network,the fine-grained classification features of images in different regions,learned by the sub-networks,are fused through a well-designed aggregation mechanism.These resulting fused features are subsequently fed into a classification module composed of fully connected layers,which is responsible for classifying both cover and stego images.This process effectively improves the overall detection performance.We carry out extensive experiments on BOSSBase-1.01 and BOWs2,two datasets that are widely utilized in the fields of steganography and steganalysis.The results show that the CES-Net achieves stateof-the-art detection accuracy for various mainstream steganography algorithms in both spatial and JPEG images,outperforming existing methods.Specifically,CES-Net improves the detection accuracy by 1.27%to 25.61%,2.1%to 21.73%,and 1.69%to 23.46%,respectively,for spatial domain algorithms(WOW,S-UNIWARD,and HILL)across different payload conditions.Furthermore,CES-Net achieves a maximum improvement of 2.34%and 2.06%in detection accuracy for two quality factors(QF=75 and QF=85)respectively when detecting JUNIWARD adaptive steganography in JPEG images.
作者
张倩倩
李浩
张祎
马媛媛
罗向阳
ZHANG Qian-Qian;LI Hao;ZHANG Yi;MA Yuan-Yuan;LUO Xiang-Yang(Key Laboratory of Cyberspace Situation Awareness of Henan Province,Information Engineering University,Zhengzhou 450001;Department of Computer and Information Engineering,Henan Normal University,Xinxiang,Henan 453007;Engineering Lab of Intelligence Business&Internet of Things,Henan Normal University,Xinxiang,Henan 453007)
出处
《计算机学报》
北大核心
2025年第6期1305-1326,共22页
Chinese Journal of Computers
基金
河南省优秀青年科学基金(252300421233,222300420058)
国家自然科学基金(U23A20305,62172435,62202495)
国家重点研发计划(2022YFB3102900)
中原学者项目(254000510007)
河南省重点研发专项基金(No.221111321200)资助。
关键词
隐写分析
隐写
孪生网络
交叉注意力机制
信息隐藏
steganalysis
steganography
siamese neural network
cross-attention
information hiding
