重庆理工大学学报2025，Vol.39Issue(11)：132-141,10.DOI:10.3969/j.issn.1674-8425(z).2025.06.016
结合残差去噪网络和多尺度深度卷积的JPEG隐写分析
JPEG steganalysis combined with residual denoising network and multi-scale deep convolution
摘要
Abstract
Contemporary steganography techniques,introducing weak noise signals into images,markedly reduce the perceptibility of the steganographic signals,thereby complicating their extraction.Moreover,existing steganalysis methods in the JPEG domain often struggle to accurately distinguish between local and global characteristics when constructing detection models.To address the problem,this paper introduces a JPEG steganalysis model that integrates residual denoising networks and multi-scale deep convolution,named DMNet.First,to enhance the steganographic signal and reduce image content interference,an improved Denoising Convolutional Neural Network is utilized to thoroughly mine noise information and improve the design of the noise reduction module.This module leverages features across various levels to capture a more comprehensive set of steganographic signals.Then,a multi-receptive field module is designed,incorporating multi-scale deep convolution to expand the receptive field,enriching the capture of local and global information and enhancing the recognition of steganographic features.Next,to eliminate redundant features,a dimension reduction module is introduced,employing average pooling for feature dimensionality reduction.Finally,in conjunction with a channel attention mechanism,the model adaptively assigns weights to multi-scale features,facilitating more precise extraction and identification of steganographic features.Results show DMNet markedly outperforms SRNet,J-XuNet,and WangNet in detection accuracy while employing steganographic method such as J-UNIWARD and UERD,with a maximum improvement of 36.14％.Furthermore,DMNet exhibits robust generalization capabilities,especially in scenarios where a mismatch exists between training and testing datasets.关键词
JPEG隐写分析/DnCNN去噪网络/多尺度特征提取/通道注意力Key words
JPEG steganalysis/DnCNN denoising network/multi-scale feature extraction/channel atten-tion分类
信息技术与安全科学引用本文复制引用
宋俊芳,王方馨,雷善中,冯飞扬..结合残差去噪网络和多尺度深度卷积的JPEG隐写分析[J].重庆理工大学学报,2025,39(11):132-141,10.基金项目
国家自然科学基金项目(62263028) （62263028）
西藏自治区自然科学基金项目(XZ202301ZR0042G) （XZ202301ZR0042G）
西藏民族大学校内科研项目(24MDY06) （24MDY06）
西藏民族大学研究生科研创新与实践项目(Y2024055,Y2024056) （Y2024055,Y2024056）
