基于多级可逆神经网络的大容量裁剪稳健型图像隐写技术

doi:10.11996/JG.j.2095-302X.2023061149

图学学报 ›› 2023, Vol. 44 ›› Issue (6): 1149-1161.DOI: 10.11996/JG.j.2095-302X.2023061149

• 图像处理与计算机视觉 • 上一篇下一篇

基于多级可逆神经网络的大容量裁剪稳健型图像隐写技术

李泓萱¹^,²(), 张松洋³, 任博²^,⁴()

1.南开大学网络空间安全学院，天津 300381
2.天津市媒体计算技术工程研究中心，天津 300381
3.中国民航信息网络股份有限公司，北京 101318
4.南开大学计算机学院，天津 300381

收稿日期:2023-07-17 接受日期:2023-09-05 出版日期:2023-12-31 发布日期:2023-12-17
通讯作者: 任博(1987-)，男，副教授，博士。主要研究方向为计算机图形学、计算机视觉等。E-mail：rb@nankai.edu.cn
作者简介:
李泓萱(1999-)，女，硕士研究生。主要研究方向为数字图像处理和计算机视觉。E-mail：hxli@mail.nankai.edu.cn
基金资助:
中央高校基础研究经费项目(63233080)

High-capacity clipped robust image steganography based on multilevel invertible neural networks

LI Hong-xuan¹^,²(), ZHANG Song-yang³, REN Bo²^,⁴()

1. College of Cyber Science, Nankai University, Tianjin 300381, China
2. Tianjin Media Computing Center, Tianjin 300381, China
3. TravelSky Technology Limited, Beijing 101318, China
4. College of Computer Science, Nankai University, Tianjin 300381, China

Received:2023-07-17 Accepted:2023-09-05 Online:2023-12-31 Published:2023-12-17
Contact: REN Bo (1987-), associate professor, Ph.D. His main research interests cover computer graphics, computer vision, etc. E-mail：rb@nankai.edu.cn
About author:
LI Hong-xuan (1999-), master student. Her main research interests cover image processing and computer vision.
E-mail：hxli@mail.nankai.edu.cn
Supported by:
The Fundamental Research Funds for the Central Universities(63233080)

摘要/Abstract

摘要：

图像隐写技术是将秘密信息嵌入到载体图像中，以保护信息的机密性，并确保不被观察者察觉。然而，在传输过程中，由于分辨率限制，载密图像的边缘区域容易受到裁剪。因此，如何从边缘区域缺失的载密图像中恢复出有效的连续隐藏信息是一个值得研究的问题。同时，图像隐写技术的另一个挑战是如何在不被检测到的情况下增加信息的有效载荷容量。为了解决上述问题，提出了一种数据驱动的图像隐写算法方案。采用了一种大容量、裁剪稳健的多级双向映射的可逆隐写网络，能够从边缘破损的载密图像中尽可能完整地恢复出连续的秘密图像。此外，算法具有高度的灵活性，可以通过多层级联中改变图像分支的通道数量实现不同规格的大尺寸图像隐写。实验表明，在各种公开数据集上生成的载密图像的视觉隐蔽性、质量度量指标和裁剪恢复能力方面显著优于其他方法。

关键词: 计算机视觉, 大容量图像隐写技术, 多级可逆神经网络, 嵌套模块级联架构, 图像裁剪稳健性

Abstract:

Image steganography aims to safeguard information confidentiality by embedding secret information into carrier images while evading detection by observers. However, during the transmission, the edges of the carrier images are often prone to cropping due to resolution limitations, making it challenging to recover continuous hidden information from the edge-missing carrier images. Another challenge in image steganography is how to enhance the effective payload capacity without being detected. To address these challenges, we proposed a data-driven image steganography algorithm that employed a high-capacity and clipped robust multilevel invertible steganography network (CR-MISN). This network had the capability to recover the continuous secret images as fully as possible from carrier images with damaged edges. Furthermore, the algorithm exhibited a high degree of flexibility, allowing for the steganography of large-sized images with different specifications by altering channel numbers in the multilevel cascading of image branches. Experimental results demonstrated that the proposed method outperformed other state-of-the-art methods in terms of visual imperceptibility, quality metrics, and cropping recovery on various public datasets.

Key words: computer vision, high-capacity image steganography, multilevel invertible neural networks, nested module cascading architecture, image clipping robustness

中图分类号:

TP391

李泓萱, 张松洋, 任博. 基于多级可逆神经网络的大容量裁剪稳健型图像隐写技术[J]. 图学学报, 2023, 44(6): 1149-1161.

LI Hong-xuan, ZHANG Song-yang, REN Bo. High-capacity clipped robust image steganography based on multilevel invertible neural networks[J]. Journal of Graphics, 2023, 44(6): 1149-1161.

图/表 13

图1 裁剪稳健型的大容量图像隐写任务流程示意图

Fig. 1 Schematic diagram of the workflow for robust and high-capacity image steganography

图2 基于可逆神经网络的嵌套模块级联架构CR-MISN示意图(输入为载体图像和大尺寸秘密图像，中间生成与载体图像视觉一致的载密图像，输出为恢复的高质量载体图像、秘密图像和中间图像)

Fig. 2 Schematic diagram of the nested module cascading architecture CR-MISN based on reversible neural networks (The input consists of a carrier image and a large-sized secret image, which generates a visually consistent stego image. The output includes the recovered high-quality carrier image, secret image, and intermediate image)

图3 嵌入-提取块内部结构图

Fig. 3 Internal structure diagram of the embedding-revealing block

图4 降采样模块与局部分块转换器示意图

Fig. 4 Schematic diagram of the downsampling module and local block transformer

表1 DIV2K[26]和Flickr2K[27]数据集上n倍载荷容量的载密图像与恢复图像PSNR，SSIM，LPIPS指标对比

Table 1 Comparison of PSNR, SSIM, LPIPS metrics for stego images and recovered images with n times payload capacity on the DIV2K[26] and Flickr2K[27] datasets

方法(n倍容量)	载密图像			秘密图像
方法(n倍容量)	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
IICNet (n=2)	36.78	0.945	0.002 2	43.02	0.992	0.002 30
ISN (n=2)	36.84	0.948	0.003 5	33.53	0.934	0.018 70
CR-MISN (n=2)	38.70	0.960	0.002 1	41.42	0.987	0.006 60
IICNet (n=3)	35.16	0.923	0.008 6	36.22	0.965	0.014 10
ISN (n=3)	34.59	0.918	0.007 1	29.63	0.867	0.052 10
CR-MISN (n=3)	36.61	0.926	0.005 3	36.37	0.968	0.012 80
IICNet (n=5)	27.95	0.747	0.010 1	34.09	0.941	0.034 50
ISN (n=5)	31.98	0.887	0.031 5	29.63	0.867	0.149 40
CR-MISN (n=5)	34.30	0.921	0.008 1	32.26	0.935	0.040 30
IICNet (n=7)	25.87	0.725	0.172 5	31.55	0.905	0.087 84
ISN (n=7)	30.36	0.862	0.097 5	26.72	0.797	0.299 54
CR-MISN (n=7)	32.64	0.917	0.028 9	30.27	0.902	0.077 89

图5 2倍、3倍载荷容量的载密图像与恢复的秘密图像的视觉比较((a) 2倍载荷容量实验结果；(b) 3倍载荷容量实验结果)

Fig. 5 Visual comparison of carrier images with 2x, 3x payload capacity and recovered secret images ((a) Experimental results of 2 times load capacity; (b) Experimental results of 3 times load capacity)

图6 5倍、7倍载荷容量的载密图像与恢复的隐藏图像的视觉比较，逐像素误差放大20倍生成误差图((a) 5倍载荷容量实验结果；(b) 7倍载荷容量实验结果)

Fig. 6 Visual comparison of carrier images with 5x, 7x payload capacity and recovered secret images ((a) Experimental results of 5 times load capacity; (b) Experimental results of 7 times load capacity)

图7 隐藏水印信息的裁剪恢复结果的视觉比较

Fig. 7 Visual comparison of cropping recovery results for hidden watermark information

表2 DIV2K[26]和Flickr2K[27]数据集上不同缺失比例的载密图像中恢复的秘密图像的指标对比

Table 2 Comparison of metrics for recovered secret images from steganographic images with different proportions of cropping attacks on DIV2K[26] and Flickr2K[27] datasets

区域缺失比例(α)	PSNR	SSIM	LPIPS
IICNet (α= 20%)	15.363	0.318	0.272
ISN (α= 20%)	15.449	0.264	0.413
CR-MISN (α= 20%)	29.539	0.921	0.047
IICNet (α= 40%)	13.953	0.288	0.406
ISN (α= 40%)	14.027	0.246	0.499
CR-MISN (α= 40%)	29.367	0.920	0.048
IICNet (α= 60%)	13.096	0.264	0.497
ISN (α= 60%)	13.210	0.228	0.557
CR-MISN (α= 60%)	27.763	0.911	0.058

图8 对载密图像进行不同方式和比例的裁剪攻击后，恢复的秘密图像的视觉比较

Fig. 8 Visual comparison of recovered secret images after clipping the loaded images in different ways and proportions

表3 DIV2K[26]和Flickr2K[27]数据集上基于载密图像损失权重的消融实验指标对比

Table 3 Comparison of metrics for ablation experiments based on loss weights of steganographic images on DIV2K[26] and Flickr2K[27] datasets

损失权重(λ_S)	载密图像			秘密图像
损失权重(λ_S)	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
λ_S= 16.0	26.78	0.804	0.172 9	33.57	0.949	0.021 8
λ_S= 32.0	28.51	0.861	0.135 1	32.97	0.943	0.028 9
λ_S= 70.0	34.30	0.921	0.008 1	32.26	0.935	0.040 3

图9 不同损失函数权重λS或不同分块比例系数Kb时，载密图像与恢复的隐藏图像的视觉比较((a)载密图像损失权重消融实验结果；(b)分块比例系数消融实验结果)

Fig. 9 Visual comparison of steganographic images and recovered secret images with different loss function weights or different block proportion coefficients ((a) Loss weight ablation results of dense image; (b) Block ratio coefficient ablation results)

表4 DIV2K[26]和Flickr2K[27]数据集上基于分块比例系数的消融实验指标对比

Table 4 Comparison of metrics for ablation experiments based on block proportion coefficients on DIV2K[26] and Flickr2K[27] datasets

损失权重(K_b)	载密图像			秘密图像
损失权重(K_b)	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
K_b= 6	32.75	0.891	0.033 9	31.07	0.908	0.067 2
K_b= 12	33.36	0.916	0.027 1	31.13	0.912	0.059 8
K_b= 72	34.30	0.921	0.008 1	32.26	0.935	0.040 3

参考文献 49

[1]	PROVOS N, HONEYMAN P. Hide and seek: an introduction to steganography[J]. IEEE Security & Privacy, 2003, 1(3): 32-44.
[2]	葛秀慧, 胡爱华, 田浩, 等. 隐写术的研究与应用[J]. 计算机应用与软件, 2007, 24(11): 57-60.
	GE X H, HU A H, TIAN H, et al. The research and application of steganography[J]. Computer Applications and Software, 2007, 24(11): 57-60 (in Chinese).
[3]	KESSLER G. An overview of cryptographic methods[EB/OL]. [2023-01-12]. https://www.google.com/books?hl=zh-CN&lr=&id=0xEIEAAAQBAJ&oi=fnd&pg=PA73&dq=An+overview+of+cryptographic+methods&ots=1OcJjqLhcz&sig=Pgf1Y_Tl5_Gy5aJalbT3pmNCsjU.
[4]	郑东, 赵庆兰, 张应辉. 密码学综述[J]. 西安邮电大学学报, 2013, 18(6): 1-10.
	ZHENG D, ZHAO Q L, ZHANG Y H. A brief overview on cryptography[J]. Journal of Xi'an University of Posts and Telecommunications, 2013, 18(6): 1-10 (in Chinese).
[5]	CHEDDAD A, CONDELL J, CURRAN K, et al. Digital image steganography: survey and analysis of current methods[J]. Signal Processing, 2010, 90(3): 727-752. DOI URL
[6]	付章杰, 王帆, 孙星明, 等. 基于深度学习的图像隐写方法研究[J]. 计算机学报, 2020, 43(9): 1656-1672.
	FU Z J, WANG F, SUN X M, et al. Research on steganography of digital images based on deep learning[J]. Chinese Journal of Computers, 2020, 43(9): 1656-1672 (in Chinese).
[7]	顾霞. 网络通信数据中的隐写技术研究[D]. 南京: 南京理工大学, 2006.
	GU X. Research on steganography techniques in network communication data[D]. Nanjing: Nanjing University of Science and Technology, 2006 (in Chinese).
[8]	SIVA SANKAR A, PRASAD T J, GIRIPRASAD M N. LSB based image steganography using polynomials and covert communications in open systems environment for DRM[C]// Proceedings of the International Conference & Workshop on Emerging Trends in Technology. New York: ACM, 2011: 593-597.
[9]	PRADHAN A, SAHU A K, SWAIN G, et al. Performance evaluation parameters of image steganography techniques[C]// 2016 International Conference on Research Advances in Integrated Navigation Systems. New York: IEEE Press, 2016: 1-8.
[10]	梁小萍, 何军辉, 李健乾, 等. 隐写分析: 原理、现状与展望[J]. 中山大学学报: 自然科学版, 2004, 43(6): 93-96.
	LIANG X P, HE J H, LI J Q, et al. Steganalysis—principle, actuality and prospect[J]. Acta Scientiarum Naturalium Universitatis Sunyatseni, 2004, 43(6): 93-96 (in Chinese).
[11]	王朔中, 张新鹏, 张卫明. 以数字图像为载体的隐写分析研究进展[J]. 计算机学报, 2009, 32(7): 1247-1263.
	WANG S Z, ZHANG X P, ZHANG W M. Recent advances in image-based steganalysis research[J]. Chinese Journal of Computers, 2009, 32(7): 1247-1263 (in Chinese). DOI URL
[12]	SHARDA S, BUDHIRAJA S. Image steganography: a review[J]. International Journal of Emerging Technology and Advanced Engineering, 2013, 3(1): 707-710.
[13]	CHAN C K, CHENG L M. Hiding data in images by simple LSB substitution[J]. Pattern Recognition, 2004, 37(3): 469-474. DOI URL
[14]	YANG H F, SUN X M, SUN G. A high-capacity image data hiding scheme using adaptive LSB substitution[J]. Radioengineering, 2009, 18(4): 509-516.
[15]	BALUJA S. Hiding images in plain sight: deep steganography[C]// The 31st International Conference on Neural Information Processing Systems. New York: ACM, 2017: 2066-2076.
[16]	ZHU J R, KAPLAN R, JOHNSON J, et al. HiDDeN: hiding data with deep networks[C]// European Conference on Computer Vision. Cham: Springer, 2018: 682-697.
[17]	BALUJA S. Hiding images within images[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(7): 1685-1697. DOI PMID
[18]	LU S P, WANG R, ZHONG T, et al. Large-capacity image steganography based on invertible neural networks[C]// 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2021: 10811-10820.
[19]	CHENG K L, XIE Y Q, CHEN Q F. IICNet: a generic framework for reversible image conversion[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2022: 1971-1980.
[20]	GUAN Z Y, JING J P, DENG X, et al. DeepMIH: deep invertible network for multiple image hiding[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(1): 372-390. DOI URL
[21]	XU Y M, MOU C, HU Y J, et al. Robust invertible image steganography[C]// 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 7865-7874.
[22]	DINH L, KRUEGER D, BENGIO Y. NICE: non-linear independent components estimation[EB/OL]. [2023-02-12]. https://arxiv.org/abs/1410.8516.
[23]	DINH L, SOHL-DICKSTEIN J, BENGIO S. Density estimation using RealNVP[EB/OL]. [2023-02-12]. https://arxiv.org/abs/1605.08803.
[24]	KINGMA D P, DHARIWAL P. Glow: generative flow with invertible 1x1 convolutions[EB/OL]. [2023-02-12]. https://arxiv.org/abs/1807.03039.
[25]	XIAO M Q, ZHENG S X, LIU C, et al. Invertible image rescaling[C]// European Conference on Computer Vision. Cham: Springer, 2020: 126-144.
[26]	AGUSTSSON E, TIMOFTE R. NTIRE 2017 challenge on single image super-resolution: dataset and study[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops. New York: IEEE Press, 2017: 1122-1131.
[27]	LIM B, SON S, KIM H, et al. Enhanced deep residual networks for single image super-resolution[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops. New York: IEEE Press, 2017: 1132-1140.
[28]	NEETA D, SNEHAL K, JACOBS D. Implementation of LSB steganography and its evaluation for various bits[C]// 2006 1st International Conference on Digital Information Management. New York: IEEE Press, 2007: 173-178.
[29]	PAN F, LI J, YANG X Y. Image steganography method based on PVD and modulus function[C]// 2011 International Conference on Electronics, Communications and Control. New York: IEEE Press, 2011: 282-284.
[30]	NGUYEN B C, YOON S M, LEE H K. Multi bit plane image steganography[M]// Digital Watermarking. Heidelberg: Springer, 2006: 61-70.
[31]	PAN Z, HU S, MA X, et al. Reversible data hiding based on local histogram shifting with multilayer embedding[J]. Journal of Visual Communication and Image Representation, 2015, 31: 64-74. DOI URL
[32]	IMAIZUMI S, OZAWA K. Multibit embedding algorithm for steganography of palette-based images[C]// Pacific-Rim Symposium on Image and Video Technology. Heidelberg: Springer, 2014: 99-110.
[33]	PO-YUEH P Y, LIN H J. A DWT based approach for image steganography[J]. International Journal of Applied Science and Engineering, 2006, 4(3): 275-290.
[34]	KAUR B, KAUR A, SINGH J. Steganographic approach for hiding image in DCT domain[J]. International Journal of Advances in Engineering & Technology, 2011, 1(3): 72.
[35]	LIAO X, YIN J J, CHEN M L, et al. Adaptive payload distribution in multiple images steganography based on image texture features[J]. IEEE Transactions on Dependable and Secure Computing, 2022, 19(2): 897-911.
[36]	ROY R, SARKAR A, CHANGDER S. Chaos based edge adaptive image steganography[J]. Procedia Technology, 2013, 10: 138-146. DOI URL
[37]	YANG J H, RUAN D Y, HUANG J W, et al. An embedding cost learning framework using GAN[J]. IEEE Transactions on Information Forensics and Security, 2020, 15: 839-851. DOI URL
[38]	TANG W X, LI B, TAN S Q, et al. CNN-based adversarial embedding for image steganography[J]. IEEE Transactions on Information Forensics and Security, 2019, 14(8): 2074-2087. DOI URL
[39]	ZHU J R, KAPLAN R, JOHNSON J, et al. HiDDeN: hiding data with deep networks[C]// European Conference on Computer Vision. Cham: Springer, 2018: 682-697.
[40]	GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial networks[J]. Communications of the ACM, 2020, 63(11): 139-144. DOI URL
[41]	JING J P, DENG X, XU M, et al. HiNet: deep image hiding by invertible network[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2022: 4713-4722.
[42]	ARDIZZONE L, LÜTH C, KRUSE J, et al. Guided image generation with conditional invertible neural networks[EB/OL]. [2023-02-12]. https://arxiv.org/abs/1907.02392.
[43]	VAN DER OUDERAA T F A, WORRALL D E. Reversible GANs for memory-efficient image-to-image translation[C]// 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 4715-4723.
[44]	ZHU X B, LI Z Z, ZHANG X Y, et al. Residual invertible spatio-temporal network for video super-resolution[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2019, 33(1): 5981-5988. DOI URL
[45]	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2017: 2261-2269.
[46]	PASZKE A, GROSS S, MASSA F, et al. PyTorch: an imperative style, high-performance deep learning library[EB/OL]. [2023-02-12]. https://arxiv.org/abs/1912.01703.
[47]	KINGMA D P, BA J. Adam: a method for stochastic optimization[EB/OL]. [2023-02-12]. https://arxiv.org/abs/1412.6980.
[48]	WANG Z, BOVIK A C, SHEIKH H R, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE Transactions on Image Processing: a Publication of the IEEE Signal Processing Society, 2004, 13(4): 600-612. DOI URL
[49]	ZHANG R, ISOLA P, EFROS A A, et al. The unreasonable effectiveness of deep features as a perceptual metric[C]// 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2018: 586-595.

基于多级可逆神经网络的大容量裁剪稳健型图像隐写技术

High-capacity clipped robust image steganography based on multilevel invertible neural networks

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 13

参考文献 49

相关文章 15

编辑推荐

Metrics

本文评价

[1]	王江安, 黄乐, 庞大为, 秦林珍, 梁温茜. 基于自适应聚合循环递归的稠密点云重建网络 [J]. 图学学报, 2024, 45(1): 230-239.
[2]	毕春艳, 刘越. 基于深度学习的视频人体动作识别综述[J]. 图学学报, 2023, 44(4): 625-639.
[3]	杨柳, 吴晓群. 基于深度学习的三维形状补全研究综述[J]. 图学学报, 2023, 44(2): 201-215.
[4]	王江安, 庞大为, 黄乐, 秦林珍. 基于多尺度特征递归卷积的稠密点云重建网络 [J]. 图学学报, 2022, 43(5): 875-883.
[5]	梁振宇, 华嘉皓, 陈浩龙, 邓逸川. 基于计算机视觉的建筑施工期临时结构损伤识别方法 [J]. 图学学报, 2022, 43(4): 608-615.
[6]	熊琛, 陈立斌, 李林泽, 许镇, 赵杨平. 基于计算机视觉与 BIM 的裂缝可视化管理方法[J]. 图学学报, 2022, 43(4): 721-728.
[7]	高铭, 张荷花, 张庭瑞, 张轩铭. 基于深度学习的公共建筑像素施工图空间识别[J]. 图学学报, 2022, 43(2): 189-196.
[8]	段锐, 邓晖, 邓逸川. ICT 支持的塔吊安全管理框架—— 回顾与展望[J]. 图学学报, 2022, 43(1): 11-20.
[9]	张宇航 , 苏成 , 邓逸川 , . 基于欧拉运动放大算法的桥梁索力测试方法研究[J]. 图学学报, 2021, 42(6): 941-947.
[10]	高文婷, 刘越. 面向移动增强现实的实时深度学习目标检测方法综述[J]. 图学学报, 2021, 42(4): 525-534.
[11]	黄凯奇, 赵鑫, 李乔哲, 胡世宇 . 视觉图灵：从人机对抗看计算机视觉下一步发展[J]. 图学学报, 2021, 42(3): 339-348.
[12]	沙浩, 刘越. 基于深度学习的图像本征属性预测方法综述[J]. 图学学报, 2021, 42(3): 385-397.
[13]	姚翰, 殷雪峰, 李童, 张肇轩, 杨鑫, 尹宝才. 基于多任务模型的深度预测算法研究[J]. 图学学报, 2021, 42(3): 446-453.
[14]	刘昶, 徐超远, 张鑫, 薛磊. 液晶字符识别的 CNN 和 SVM 组合分类器[J]. 图学学报, 2021, 42(1): 15-22.
[15]	陈炎，杨丽丽，王振鹏. 双目视觉的匹配算法综述[J]. 图学学报, 2020, 41(5): 702-708.