基于改进U-Net和混合注意力机制的高质量全尺寸图像隐写方法

doi:10.3969/j.issn.1671-1122.2024.07.007

摘要/Abstract

摘要：

图像隐写术是一种将秘密信息隐藏在图像中以防止被发现的技术。当前的图像隐写模型存在图像生成质量差和抗隐写分析能力弱等问题。混合注意力机制不仅可以抑制无意义的通道信息，避免其在含密图像上产生伪影，而且可以根据载体图像中不同位置的重要程度分配不同的权重，为秘密信息寻找更合适的隐藏区域。基于上述特点，文章提出了基于U-Net和混合注意力机制的高质量全尺寸图像隐写方法。文章所提方法的模型由编码器、提取器和判别器三个子网组成。编码器采用改进的U-Net结构和混合注意力机制模块进行设计；提取器采用卷积神经网络和混合注意力机制模块进行设计；判别器则用于增强模型的安全性。实验结果表明，该方法能够将尺寸为256×256的彩色秘密图像完全隐藏在同尺寸的彩色载体图像中，在不降低隐写容量的情况下实现高质量的图像隐写。该方法在ImageNet、COCO、DIV2K三个数据集上都展现了良好的视觉质量和隐藏容量，在ImageNet数据集上，PSNR值最高可达40.143 dB（载体图像与含密图像）和42.082 dB（秘密图像与重构的秘密图像），同时还能够提高模型的抗隐写分析能力。

关键词: 图像隐写, 混合注意力机制, U-Net

Abstract:

Image steganography is a technique for hiding secret information within images to prevent detection. Current image steganography models face issues, such as poor image generation quality and weak resistance to steganalysis. The hybrid attention mechanism can suppress meaningless channel information to avoid artifacts in the stego image and assign different weights based on the importance of different positions in the cover image, thereby finding more suitable hiding areas for the secret information. Based on these features, this paper proposed a high-quality, full-size image steganography method based on U-Net and hybrid attention mechanisms. The model comprised three subnetworks: an encoder, an extractor, and a discriminator. The encoder was designed using an improved U-Net structure and a hybrid attention mechanism module; the extractor was designed using convolutional neural networks and a hybrid attention mechanism module; the discriminator enhanced the model’s security. Experimental results show that this method can completely hide a 256×256 color secret image within a cover image of the same size, achieving high-quality image steganography without reducing the steganographic capacity. This method demonstrates good visual quality and hiding capacity on the ImageNet, COCO, and DIV2K datasets. On the ImageNet dataset, the PSNR values can reach up to 40.143 dB (cover image and stego image) and 42.082 dB (secret image and reconstructed secret image), while also improving the model’s resistance to steganalysis.

Key words: image steganography, hybrid attention mechanism, U-Net

中图分类号:

TP309

董云云, 朱玉玲, 姚绍文. 基于改进U-Net和混合注意力机制的高质量全尺寸图像隐写方法[J]. 信息网络安全, 2024, 24(7): 1050-1061.

DONG Yunyun, ZHU Yuling, YAO Shaowen. High-Quality Full-Size Image Steganography Method Based on Improved U-Net and Hybrid Attention Mechanism[J]. Netinfo Security, 2024, 24(7): 1050-1061.

图/表 16

图1

图2

图3

表1

图4

表2

图5

表3

图6

图7

图8

图9

表4

表5

表6

表7

参考文献 25

[1]	MANDAL P C, MUKHERJEE I, PAUL G, et al. Digital Image Steganography: A Literature Survey[J]. Information Sciences, 2022, 609: 1451-1488.
[2]	XIA Qiang, HE Peisong, LUO Jie, et al. An Efficient Enhancement Algorithm of Cover Image Based on Universal Adversarial Noise[J]. Netinfo Security, 2022, 22(2): 64-75.
	夏强, 何沛松, 罗杰, 等. 基于普遍对抗噪声的高效载体图像增强算法[J]. 信息网络安全, 2022, 22(2): 64-75.
[3]	ZHANG Hongjuan, ZHU Chenming. Novel LSB Steganography Algorithm of Against Statistical Analysis[J]. Computer Engineering, 2008, 34(23): 144-146. doi: 10.3969/j.issn.1000-3428.2008.23.052
[4]	LI Xiaolong, YANG Bin, CHENG Daofang, et al. A Generalization of LSB Matching[J]. IEEE Signal Processing Letters, 2009, 16(2): 69-72.
[5]	HOLUB V, FRIDRICH J. Designing Steganographic Distortion Using Directional Filters[C]// IEEE. 2012 IEEE International Workshop on Information Forensics and Security (WIFS). New York: IEEE, 2012: 234-239.
[6]	PEVNY T, FILLER T, BAS P. Using High-Dimensional Image Models to Perform Highly Undetectable Steganography[C]// Springer. 12th Information Hiding Conference. Heidelberg: Springer, 2010: 161-177.
[7]	LI Bin, WANG Ming, HUANG Jiwu, et al. A New Cost Function for Spatial Image Steganography[C]// IEEE. 2014 IEEE International Conference on Image Processing (ICIP). New York: IEEE, 2014: 4206-4210.
[8]	HOLUB V, FRIDRIVH J, DENEMARK T. Universal Distortion Function for Steganography in an Arbitrary Domain[J]. EURASIP Journal on Information Security, 2014, 2014: 1-13.
[9]	CHAUMONT M. Deep Learning in Steganography and Steganalysis[M]. New York: Academic Press, 2020.
[10]	YU Chong. Attention Based Data Hiding with Generative Adversarial Networks[C]// AAAI. 34th AAAI Conference on Artificial Intelligence. Menlo Park: AAAI, 2020: 1120-1128.
[11]	TAN Jingxuan, LIAO Xin, LIU Jiate, et al. Channel Attention Image Steganography with Generative Adversarial Networks[J]. IEEE Transactions on Network Science and Engineering, 2021, 9(2): 888-903.
[12]	ZHANG Le, LU Yao, LI Jinxing, et al. Deep Adaptive Hiding Network for Image Hiding Using Attentive Frequency Extraction and Gradual Depth Extraction[J]. Neural Computing and Applications, 2023, 35(15):10909-10927.
[13]	GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative Adversarial Networks[J]. Communications of the ACM, 2020, 63(11): 139-144.
[14]	VOLKHONSKIY D, NAZAROV I, BURNAEV E. Steganographic Generative Adversarial Networks[C]// ICMV. Twelfth International Conference on Machine Vision (ICMV 2019). Bellingham: SPIE, 2020: 991-1005.
[15]	WANG Yaojie, NIU Ke, YANG Xiaoyuan. Image Steganography Scheme Based on GANs[J]. Netinfo Security, 2019, 19(5): 54-60.
	王耀杰, 钮可, 杨晓元. 基于生成对抗网络的信息隐藏方案[J]. 信息网络安全, 2019, 19(5): 54-60.
[16]	LEI Yu, LIU Jia, LI Jun, et al. Research and Implementation of a Image Steganography Method Based on Conditional Generative Adversarial Networks[J]. Netinfo Security, 2021, 21(11): 48-57.
	雷雨, 刘佳, 李军, 等. 一种基于条件生成对抗网络的图像隐写方法研究与实现[J]. 信息网络安全, 2021, 21(11): 48-57.
[17]	BALUJA S. Hiding Images within Images[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 42(7): 1685-1697.
[18]	XIONG Jinbo, MA Rong, ZHANG Yuanyuan, et al. Image Information Hiding Method and Implementation for Social Network[J]. Netinfo Security, 2017, 17(3): 6-8.
	熊金波, 马蓉, 张媛媛, 等. 面向社交网络的图片信息隐藏方法与实现[J]. 信息网络安全, 2017, 17(3): 6-8.
[19]	ZHU Jiren, KAPLAN R, JOHNSON J, et al. Hidden: Hiding Data with Deep Networks[C]// ECCV. 15th European Conference on Computer Vision. Heidelberg: Springer, 2018: 657-672.
[20]	WENG Xinyu, LI Yongzhi, CHI Lu, et al. High-Capacity Convolutional Video Steganography with Temporal Residual Modeling[C]// ACM. Proceedings of the 2019 on International Conference on Multimedia Retrieval. New York: ACM, 2019: 87-95.
[21]	RUSSAKOVSKY O, DENG Jia, SU Hao, et al. Imagenet Large Scale Visual Recognition Challenge[J]. International Journal of Computer Vision, 2015, 115: 211-252.
[22]	LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO: Common Objects in Context[C]// ECCV. 13th European Conference on Computer Vision. Heidelberg: Springer, 2014: 740-755.
[23]	AGUSTSSON E, TIMOFTE R. Ntire 2017 Challenge on Single Image Super-Resolution: Dataset and Study[C]// IEEE. 30th IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). New York: IEEE, 2017: 126-135.
[24]	RUSTAD S, ANDONO P N, SHIDIK G F. Digital Image Steganography Survey and Investigation (Goal, Assessment, Method, Development, and Dataset)[J]. Signal Processing, 2023, 27(5): 985-1024.
[25]	AGRAWAL R, AHUJA K. CSIS: Compressed Sensing-Based Enhanced-Embedding Capacity Image Steganography Scheme[J]. IET Image Processing, 2021, 15(9): 1909-1925.

输入尺寸 (H∙W∙C)	网络结构	输出尺寸(H∙W∙C)
256∙256∙6	Conv1(4∙4,2,1)+BN+LeakyReLU	128∙128∙64
	Conv2(4∙4,2,1)+BN+HAM+LeakyReLU	64∙64∙128
	Conv3(4∙4,2,1)+BN+HAM+LeakyReLU	32∙32∙256
	Conv4(4∙4,2,1)+BN+HAM+LeakyReLU	16∙16∙512
	Conv5(4∙4,2,1)+BN+HAM+LeakyReLU	8∙8∙512
	Conv6(4∙4,2,1)+BN+HAM+LeakyReLU	4∙4∙512
	Conv7(4∙4,2,1)+BN+LeakyReLU	2∙2∙512
	DeConv1[ConvTranspose2d(4∙4,2,1)+BN+ReLU]	4∙4∙512
	DeConv2[ConvTranspose2d(4∙4,2,1)+BN+HAM+ReLU]	8∙8∙512
	DeConv3[ConvTranspose2d(4∙4,2,1)+BN+HAM+ReLU]	16∙16∙512
	DeConv4[ConvTranspose2d(4∙4,2,1)+BN+HAM+ReLU]	32∙32∙256
	DeConv5[ConvTranspose2d(4∙4,2,1)+BN+HAM+ReLU]	64∙64∙128
	DeConv6[ConvTranspose2d(4∙4,2,1)+BN+HAM+ReLU]	128∙128∙64
	DeConv7[ConvTranspose2d(4∙4,2,1)+Sigmoid]	256∙256∙3

输入尺寸(H∙W∙C)	网络结构	输出尺寸(H∙W∙C)
256∙256∙3	Conv1(3∙3,1,1)+BN+Relu	256∙256∙64
	Conv2(3∙3,1,1)+BN+Relu+HAM	256∙256∙128
	Conv3(3∙3,1,1)+BN+Relu	256∙256∙256
	Conv4(3∙3,1,1)+BN+Relu+HAM	256∙256∙512
	Conv5(3∙3,1,1)+BN+Relu	256∙256∙256
	Conv6(3∙3,1,1)+BN+Relu	256∙256∙128
	Conv7(3∙3,1,1)+BN+Relu	256∙256∙16
	Conv8(3∙3,1,1)+BN+Tanh	256∙256∙3

输入尺寸(H∙W∙C)	网络结构	输出尺寸(H∙W∙C)
256∙256∙3	Conv1(3∙3,1,1)+BN+Relu	256∙256∙64
	Conv2(3∙3,1,1)+BN+Relu	256∙256∙64
	Conv3(3∙3,1,1)+BN+Relu	256∙256∙64
	AdaptiveAvgPool2d	64
	Linear	1

方法	载体图像/含密图像				秘密图像/重构的秘密图像
方法	PSNR↑	SSIM↑	MAE↓	RMSE↓	PSNR↑	SSIM↑	MAE↓	RMSE↓
HiDDeN^[19]	37.351	0.948	2.655	3.490	36.317	0.954	3.061	3.921
文献[20] 方法	38.431	0.947	2.375	3.079	38.511	0.973	2.279	3.048
文献[17] 方法	38.978	0.975	2.288	2.952	38.001	0.970	2.427	3.232
DAH-Net^[12]	38.184	0.969	2.404	3.142	38.035	0.921	1.819	3.197
本文模型	40.143	0.962	1.893	2.515	42.082	0.986	1.494	2.011

方法	载体图像/含密图像				秘密图像/重构的秘密图像
方法	PSNR↑	SSIM↑	MAE↓	RMSE↓	PSNR↑	SSIM↑	MAE↓	RMSE↓
HiDDeN^[19]	37.270	0.951	2.636	3.512	37.410	0.970	2.651	3.470
文献[20 方法	38.106	0.959	2.419	3.188	38.707	0.975	2.184	2.986
文献[17] 方法	38.577	0.978	2.388	3.059	38.573	0.977	2.246	3.075
DAH-Net^[12]	37.323	0.963	2.684	3.470	38.203	0.975	2.295	3.315
本文模型	39.368	0.965	2.019	2.743	41.679	0.986	1.558	2.121