DDAC：面向卷积神经网络图像隐写分析模型的特征提取方法

doi:10.11959/j.issn.1000-436x.2022089

摘要/Abstract

摘要：

针对基于卷积神经网络的图像隐写分析方法中使用人工设计的滤波器在特征提取过程中有效性低的问题，提出方向差分自适应组合（DDAC）特征提取方法。在计算中心像素与周围不同方向像素的差分后，使用 1× 1 卷积对方向差分进行线性组合。根据损失对组合参数自适应更新来构建多样化的滤波器，使获取的嵌入信息残差特征更有效。使用截断线性单元提高嵌入信息残差和图像信息残差的比率，加快模型收敛速度并提高残差特征提取能力。实验结果表明，该方法使 Ye-net、Yedroudj-net 模型的准确率在 WOW 和S-UNIWARD数据集中提高1.30%～8.21%。与固定和更新参数SRM滤波器方法相比，测试模型在不同隐写数据集中的准确率提高 0.60%～20.72%，并且训练过程更稳定。对比其他图像隐写分析模型，DDAC-net 具有更高的隐写分析效率。

关键词: 图像隐写分析, 卷积神经网络, 特征提取, 隐写分析富模型, 截断线性单元

Abstract:

To solve the problem that for image steganalysis based on convolution neural network, manual designed filter kernels were used to extract residual characteristics, but in practice, these kernels filter were not suitable for each steganography algorithm and have worse performance in application, a directional difference adaptive combination (DDAC) method was proposed.Firstly, the difference was calculated between center pixel and each directional pixel around, and 1 × 1 convolution was adopted to achieve linear combinations of directional difference.Since the combination parameters self-adaptively update according to loss function, filter kernels could be more effective in extracting diverse residual characteristics of embedding information.Secondly, truncated linear unit (TLU) was applied to raise the ratio of embedding information residual to image information residual.The model’s coveragence was accelerated and the ability of feature extraction was promoted.Experimental results indicate that substituting the proposed method could improve the accuracy of Ye-net and Yedroudj-net by 1.30%~8.21% in WOW and S-UNIWARD datasets.Compared with fix and adjustable SRM filter kernels methods, the accuracy of test model using DDAC increases 0.60%~20.72% in various datasets, and the training progress was more stable.DDAC-net was proved to be more effective in comparsion with other steganalysis model.

Key words: image steganalysis, convolution neutral network, feature extraction, rich model of steganalysis, truncated linear unit

中图分类号:

TP309

王晓丹, 李京泰, 宋亚飞. DDAC：面向卷积神经网络图像隐写分析模型的特征提取方法[J]. 通信学报, 2022, 43(5): 68-81.

Xiaodan WANG, Jingtai LI, Yafei SONG. DDAC: a feature extraction method for model of image steganalysis based on convolutional neural network[J]. Journal on Communications, 2022, 43(5): 68-81.

图/表 19

图1

图2

图3

图4

图5

图6

图7

表1

图8

图9

图10

表2

图11

表3

表4

表5

表6

表7

表8

参考文献 33

[1]	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.
[2]	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.
[3]	陈君夫, 付章杰, 张卫明 ,等. 基于深度学习的图像隐写分析综述[J]. 软件学报, 2021,32(2): 551-578.
	CHEN J F , FU Z J , ZHANG W M ,et al. Review of image steganalysis based on deep learning[J]. Journal of Software, 2021,32(2): 551-578.
[4]	LECUN Y , BENGIO Y , HINTON G . Deep learning[J]. Nature, 2015,521(7553): 436-444.
[5]	KRIZHEVSKY A , SUTSKEVER I , HINTON G E . ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017,60(6): 84-90.
[6]	HE K M , ZHANG X Y , REN S Q ,et al. Deep residual learning for image recognition[C]// Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2016: 770-778.
[7]	RONNEBERGER O , FISCHER P , BROX T . U-net:convolutional networks for biomedical image segmentation[C]// Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. Berlin:Springer, 2015: 234-241.
[8]	HE K M , ZHANG X Y , REN S Q ,et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015,37(9): 1904-1916.
[9]	GOODFELLOW I J , POUGET-ABADIE J ,, MIRZA M , et al . Generative adversarial nets[C]// Proceedings of the 27th International Conference on Neural Information Processing Systems. Cambridge:MIT Press, 2014: 2672-2680.
[10]	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.
[11]	TANG W X , TAN S Q , LI B ,et al. Automatic steganographic distortion learning using a generative adversarial network[J]. IEEE Signal Processing Letters, 2017,24(10): 1547-1551.
[12]	KRIZHEVSKY A , SUTSKEVER I , HINTON G E . ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017,60(6): 84-90.
[13]	TAN S Q , LI B . Stacked convolutional auto-encoders for steganalysis of digital images[C]// Proceedings of Signal and Information Processing Association Annual Summit and Conference (APSIPA). Piscataway:IEEE Press, 2014: 1-4.
[14]	QIAN Y L , DONG J , WANG W ,et al. Deep learning for steganalysis via convolutional neural networks[C]// Proceedings of SPIE - The International Society for Optical Engineering. Bellingham:SPIE Press, 2015: 171-180.
[15]	XU G S , WU H Z , SHI Y Q . Structural design of convolutional neural networks for steganalysis[J]. IEEE Signal Processing Letters, 2016,23(5): 708-712.
[16]	YEDROUDJ M , COMBY F , CHAUMONT M . Yedroudj-net:an efficient CNN for spatial steganalysis[C]// Proceedings of 2018 IEEE International Conference on Acoustics,Speech and Signal Processing. Piscataway:IEEE Press, 2018: 2092-2096.
[17]	YE J , NI J Q , YI Y . Deep learning hierarchical representations for image steganalysis[J]. IEEE Transactions on Information Forensics and Security, 2017,12(11): 2545-2557.
[18]	FRIDRICH J , KODOVSKY J . Rich models for steganalysis of digital images[J]. IEEE Transactions on Information Forensics and Security, 2012,7(3): 868-882.
[19]	LI B , WEI W H , FERREIRA A ,et al. ReST-net:diverse activation modules and parallel subnets-based CNN for spatial image steganalysis[J]. IEEE Signal Processing Letters, 2018,25(5): 650-654.
[20]	KALMAN B L , KWASNY S C . Why tanh:choosing a sigmoidal function[C]// Proceedings of International Joint Conference on Neural Networks. Piscataway:IEEE Press, 1992: 578-581.
[21]	BOROUMAND M , CHEN M , FRIDRICH J . Deep residual network for steganalysis of digital images[J]. IEEE Transactions on Information Forensics and Security, 2019,14(5): 1181-1193.
[22]	SZEGEDY C , IOFFE S , VANHOUCKE V ,et al. Inception-v4,inception-ResNet and the impact of residual connections on learning[C]// Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence. New York:ACM Press, 2017: 4278-4284.
[23]	ZHANG R , ZHU F , LIU J Y ,et al. Depth-wise separable convolutions and multi-level pooling for an efficient spatial CNN-based steganalysis[J]. IEEE Transactions on Information Forensics and Security, 2020,15: 1138-1150.
[24]	PEVNY T , BAS P , FRIDRICH J . Steganalysis by subtractive pixel adjacency matrix[J]. IEEE Transactions on Information Forensics and Security, 2010,5(2): 215-224.
[25]	SUTSKEVER I , MARTENS J , DAHL G ,et al. On the importance of initialization and momentum in deep learning[C]// International Conference on Machine Learning. Saarland:DBLP, 2013: 1139-1147.
[26]	IOFFE S , SZEGEDY C . Batch normalization:accelerating deep network training by reducing internal covariate shift[C]// Proceedings of the 32nd International Conference on Machine Learning. Saarland:DBLP, 2015: 448-456.
[27]	NAIR V , HINTON G E . Rectified linear units improve restricted boltzmann machines[C]// Proceedings of the 27th International Conference on Machine Learning. Saarland:DBLP, 2010: 807-814.
[28]	BAS P , FILLER T , PEVNY T , . “Break our steganographic system”:the ins and outs of organizing BOSS[C]// Information Hiding. Berlin:Springer, 2011: 59-70.
[29]	KODOVSKY J , FRIDRICH J , HOLUB V . Ensemble classifiers for steganalysis of digital media[J]. IEEE Transactions on Information Forensics and Security, 2012,7(2): 432-444.
[30]	沈军, 廖鑫, 秦拯 ,等. 基于卷积神经网络的低嵌入率空域隐写分析[J]. 软件学报, 2021,32(9): 2901-2915.
	SHEN J , LIAO X , QIN Z ,et al. Spatial steganalysis of low embedding rate based on convolutional neural network[J]. Journal of Software, 2021,32(9): 2901-2915.
[31]	YOU W K , ZHANG H , ZHAO X F . A siamese CNN for image steganalysis[J]. IEEE Transactions on Information Forensics and Security, 2021,16: 291-306.
[32]	ARIVAZHAGAN S , AMRUTHA E , SYLVIA L J W ,et al. Hybrid convolutional neural network architecture driven by residual features for steganalysis of spatial steganographic algorithms[J]. Neural Computing and Applications, 2021,33(17): 11465-11485.
[33]	QIAN Y L , DONG J , WANG W ,et al. Learning and transferring representations for image steganalysis using convolutional neural network[C]// Proceedings of 2016 IEEE International Conference on Image Processing. Piscataway:IEEE Press, 2016: 2752-2756.

结构	WOW		S-UNIWARD
结构	嵌入率为0.2	嵌入率为0.4	嵌入率为0.2	嵌入率为0.4
SRM方法	0.738 3	0.844 2	0.691 3	0.803 1
结构1	0.775 9	0.870 3	0.718 3	0.845 2
结构2	0.761 5	0.845 1	0.708 3	0.835 4
结构3	0.756 9	0.857 7	0.689 2	0.823 4
结构4	0.730 8	0.832 0	0.695 1	0.820 7
结构5	0.732 7	0.846 1	0.685 0	0.810 4
结构6	0.730 5	0.831 2	0.689 7	0.808 9

方法	WOW		S-UNIWARD
方法	嵌入率为0.2	嵌入率为0.4	嵌入率为0.2	嵌入率为0.4
固定SRM滤波器	0.747 9	0.849 1	0.663 9	0.802 8
可变SRM滤波器	0.672 6	0.743 9	0.582 8	0.636 8
DDAC	0.753 9	0.872 3	0.694 7	0.844 0

模型名称	训练超参数
Ye-net	优化算法 Adadelta, 学习率为 0.4, 动量为 0.95, eps=1×10^-8, 权重缩减为 5×10^-4, 批次大小为 32, 学习率改变策略: 当训练轮次等于 100、150 时学习率乘以0.2, 训练轮数为200
Yedroudj-net	优化算法SGD, 学习率为0.01, 动量为0.95, 权重缩减y=1×10^-4, 批次大小为 16, 学习率改变策略: 当训练轮次等于100、150时学习率乘以0.1, 训练轮数为200

模型	WOW		S-UNIWARD
模型	嵌入率为0.4	嵌入率为0.2	嵌入率为0.4	嵌入率为0.2
Ye-net	0.7918	0.6749	0.7432	0.6296
Ye-net+DDAC	0.8135(+0.0217)	0.7081(+0.0332)	0.7645(+0.0213)	0.6427(+0.0131)
Yedroudj-net	0.8573	0.7495	0.7540	0.6264
Yedroudj-net+DDAC	0.8703(+0.0130)	0.7743(+0.0248)	0.8361(+0.0821)	0.7019(+0.0755)

模型	WOW		S-UNIWARD
模型	嵌入率为0.4	嵌入率为0.2	嵌入率为0.4	嵌入率为0.2
Ye-net	0.1950	0.3272	0.2882	0.3570
Ye-net+DDAC	0.2166(+0.0216)	0.2470(-0.0802)	0.2250(-0.0632)	0.3466(-0.0104)
Yedroudj-net	0.1654	0.2978	0.2454	0.4980
Yedroudj-net+DDAC	0.1068(-0.0586)	0.2372(-0.0606)	0.1476(-0.0978)	0.3042(-0.1938)