Steganographer Detection via Multiple-instance Learning Graph Convolutional Networks
-
摘要: 隱寫者檢測通過設計模型檢測在批量圖像中嵌入秘密信息進行隱蔽通信的隱寫者, 對解決非法使用隱寫術的問題具有重要意義. 本文提出一種基于多示例學習圖卷積網絡 (Multiple-instance learning graph convolutional network, MILGCN) 的隱寫者檢測算法, 將隱寫者檢測形式化為多示例學習(Multiple-instance learning, MIL) 任務. 本文中設計的共性增強圖卷積網絡(Graph convolutional network, GCN) 和注意力圖讀出模塊能夠自適應地突出示例包中正示例的模式特征, 構建有區分度的示例包表征并進行隱寫者檢測. 實驗表明, 本文設計的模型能夠對抗多種批量隱寫術和與之對應的策略.Abstract: Steganographer detection aims to solve the problem of illegal use of batch steganography by designing models to detect steganographers who embed secret information in images for covert communication. This paper proposes a novel steganographer detection algorithm called as multiple-instance learning graph convolutional network (MILGCN) to formalize steganography detection as a multiple-instance learning (MIL) task. The commonness enhancement graph convolutional network (GCN) and attention graph readout module designed in this paper can adaptively highlight the positive instance pattern and construct distinguishable bag representations for steganographer detection. Experiments show that the designed model can successfully attack a variety of batch steganography and the corresponding strategies.
-
表 1 使用的變量符號及對應說明
Table 1 The variable symbols and their corresponding descriptions
變量 含義 $B_x$ 用戶$x$對應的示例包 $x_i$ 示例包$B_x$內第$i$個示例 $m$ 當前數據集中示例包的總數量 $n$ 當前示例包中示例的總數量 $v_i$ 共性增強圖卷積模塊的第$i$個輸入示例特征 $h_i$ 對$v_i$使用$f$函數進行特征提取后得到的示例特征向量 $H$ 由示例特征向量構成的示例包的矩陣表示 $H=[h_1,\cdots,h_n]^{\rm{T}}$ $A_{ij}$ 圖中第$i$個與第$j$個示例節點之間邊的權重 $N_i$ 圖中第$i$個示例節點的所有鄰居節點 $A$ 示例包$B_x$所構成圖的鄰接矩陣 $r_i$ 進行圖卷積后$h_i$所對應的示例特征向量 $R$ 由示例特征向量構成的示例包的矩陣表示 $R=[r_1,\cdots,r_n]^{\rm{T}}$ $s_i$ 進行圖歸一化后$r_i$ 所對應的示例特征向量 $S$ 由示例特征向量構成的示例包的矩陣表示 $S=[s_1,\cdots,s_n]^{\rm{T}}$ $t_i$ 共性增強圖卷積模塊的第$i$個輸出示例特征 $T$ 由示例特征向量構成的示例包的矩陣表示 $T=[t_1,\cdots,t_n]^{\rm{T}}$ $f$ 特征提取函數 $g$ 注意力計算函數 $z_i$ 共性增強圖卷積模塊的輸出, 也是注意力讀出模塊的第$i$個輸入示例特征 $Z_x$ 由共性增強圖卷積模塊得到的示例特征向量構成的用戶$x$對應的示例包的矩陣表示 $u_x$ 用戶$x$對應的示例包的特征向量表征 $p_i$ 當前示例包中第$i$個示例對示例包表征的貢獻 ${\rho_i}$ 第$i$個示例包的預測結果 $Y_i$ 第$i$個示例包的真實標簽 $L$ 本文設計的損失函數 $L_{\rm{bag}}$ 本文設計的多示例學習分類損失 $L_{\rm{entropy}}$ 本文設計的熵正則損失 $L_{\rm{contrastive}}$ 本文設計的對比學習損失 $\lambda_1, \lambda_2, \lambda_3$ 超參數, 用于調整$L_{\rm{bag}},$ $L_{\rm{entropy}},$ $L_{\rm{contrastive}}$的權重 表 2 已知隱寫者使用相同圖像隱寫術(S-UNIWARD) 時的隱寫者檢測準確率(%), 嵌入率從0.05 bpp到0.4 bpp
Table 2 Steganography detection accuracy rate (%) when steganographers use the same image steganography (S-UNIWARD), while the embedding payload is from 0.05 bpp to 0.4 bpp
模型 嵌入率(bpp) 0.05 0.1 0.2 0.3 0.4 前沿 MDNNSD 4 54 100 100 100 XuNet_SD 2 2 71 100 100 基于GAN SSGAN_SD 0 1 1 2 4 基于GNN GAT 2 3 3 3 4 GraphSAGE 28 88 100 100 100 AGNN 24 99 100 100 100 GCN 19 96 100 100 100 SAGCN 72 100 100 100 100 基于MIL MILNN_self 15 87 100 100 100 MILNN_git 18 96 100 100 100 本文 MILGCN-MF 47 100 100 100 100 MILGCN 74 100 100 100 100 表 3 當測試階段隱寫者使用相同隱寫術(S-UNIWARD) 和分享的載密圖像數量占總圖像數量為10%到100%時, SRNet-AVG和SRNet-MILGCN的檢測成功率 (%)
Table 3 The accurate rate (%) of SRNet-AVG and SRNet-MILGCN when the number of shared secret images is from 10% to 100% of the total number of images and the steganographer uses the same steganography (S-UNIWARD) in test
方法 占比(%) 10 30 50 70 90 100 SRNet-AVG 26 100 100 100 100 100 SRNet-MILGCN 35 100 100 100 100 100 表 4 當用戶分享不同數量的圖像時, 使用MILGCN和SAGCN進行隱寫者檢測的準確率(%),嵌入率從0.05 bpp到0.4 bpp
Table 4 Steganography detection accuracy rate (%) of MILGCN and SAGCN when users share different numbers of images, while the embedding payload is from 0.05 bpp to 0.4 bpp
數量(張) 嵌入率 (bpp) 0.05 0.1 0.2 0.3 0.4 MILGCN 100 35 96 100 100 100 200 74 100 100 100 100 400 96 100 100 100 100 600 100 100 100 100 100 SAGCN 100 31 96 100 100 100 200 72 100 100 100 100 400 91 100 100 100 100 600 91 100 100 100 100 表 5 在隱寫術錯配情況下, 當分享的載密圖像數量占比5%時, MILGCN取得的隱寫者檢測準確率(%)
Table 5 Steganography detection accuracy rate (%) in the case of steganography mismatch when the number of shared secret images is 5% of the total number of images
測試隱寫術 HUGO-BD WOW HILL MiPOD 檢測準確率 6 4 5 5 表 6 訓練模型使用HILL作為隱寫術, 分享的載密圖像數量占比10%或30%, MILGCN取得的隱寫者檢測準確率(%)
Table 6 Steganography detection accuracy rate (%) when the steganography used for training is HILL and the number of shared secret images is 10% or 30% of the total number of images
載密圖像比例 測試隱寫術 HUGO-BD WOW HILL MiPOD 10% 9 6 7 4 30% 37 48 49 47 表 7 已知隱寫者使用相同圖像隱寫術(J-UNIWARD)時的隱寫者檢測準確率(%), 嵌入率從0.05 bpnzAC到0.4 bpnzAC
Table 7 Steganography detection accuracy rate (%) when steganographer use the same image steganography (J-UNIWARD) and the embedding payload is from 0.05 bpnzAC to 0.4 bpnzAC
模型 嵌入率(bpnzAC) 0.05 0.1 0.2 0.3 0.4 JRM_SD 11 17 25 31 48 PEV_SD 0 0 1 1 5 GraphSAGE 13 68 100 100 100 AGNN 13 84 100 100 100 GCN 16 88 100 100 100 SAGCN 17 92 100 100 100 MILGCN 25 92 100 100 100 表 8 當測試階段隱寫者使用nsF5或UERD等圖像隱寫術嵌入秘密信息時, 不同方法的隱寫者檢測準確率(%),嵌入率從0.05 bpnzAC到0.4 bpnzAC
Table 8 Steganography detection accurate rate (%) of different methods when steganographer uses nsF5 or UERD as image steganography in the testing phase and the embedding payload is from 0.05 bpnzAC to 0.4 bpnzAC
隱寫術 模型 嵌入率(bpnzAC) 0.05 0.1 0.2 0.3 0.4 nsF5 PEV_SD 0 1 9 52 93 GraphSAGE 21 91 100 100 100 AGNN 20 90 100 100 100 GCN 24 90 100 100 100 SAGCN 29 92 100 100 100 MILGCN 22 90 100 100 100 UERD GraphSAGE 25 91 100 100 100 AGNN 29 94 100 100 100 GCN 33 96 100 100 100 SAGCN 33 98 100 100 100 MILGCN 42 99 100 100 100 表 9 計算復雜度分析
Table 9 The analysis of computational complexity
方法名稱 批次平均
運行時間(s)單個樣本浮點
運算數(千兆次)參數量(千個) MILNN 0.001 0.003 12.92 GCN 0.830 2.480 67.97 SAGCN 2.210 7.410 67.94 MILGCN 0.020 0.070 74.18 亚洲第一网址_国产国产人精品视频69_久久久久精品视频_国产精品第九页 -
[1] Ker A D, Pevny T. A new paradigm for steganalysis via clustering. In: Proceedings of the SPIE 7880, Media Watermarking, Security, and Forensics III. Bellingham, USA: SPIE, 2011. 312?324 [2] Pevny T, Feidrich J. Merging Markov and DCT features for multi-class JPEG steganalysis. In: Proceedings of the SPIE 6505, Security, Steganography, and Watermarking of Multimedia Contents IX. Bellingham, USA: SPIE, 2007. 1?13 [3] Ker A D, Pevny T. Identifying a steganographer in realistic and heterogeneous data sets. In: Proceedings of the SPIE 8303, Media Watermarking, Security, and Forensics. Bellingham, USA: SPIE, 2012. 1?13 [4] Ker A D, Pevny T. The steganographer is the outlier: Realistic large-scale steganalysis. IEEE Transactions on Information Forensics and Security, 2014, 9(9): 1424?1435 doi: 10.1109/TIFS.2014.2336380 [5] Li F Y, Wu K, Lei J S, Wen M, Bi Z Q, Gu C H. Steganalysis over large-scale social networks with high-order joint features and clustering ensembles. IEEE Transactions on Information Forensics and Security, 2016, 11(2): 344?357 doi: 10.1109/TIFS.2015.2496910 [6] Zheng M J, Zhong S H, Wu S T, Jiang J M. Steganographer detection via deep residual network. In: Proceedings of the IEEE International Conference on Multimedia and Expo (ICME). Piscataway, USA: IEEE, 2017. 235?240 [7] Zheng M J, Zhong S H, Wu S T, Jiang J M. Steganographer detection based on multiclass dilated residual networks. In: Proceedings of the ACM on International Conference on Multimedia Retrieval. New York, USA: ACM, 2018. 300?308 [8] Zhang Z, Zheng M J, Zhong S H, Liu Y. Steganographer detection via enhancement-aware graph convolutional network. In: Proceedings of the IEEE International Conference on Multimedia and Expo (ICME). Piscataway, USA: IEEE, 2020. 1?6 [9] Zhang Z, Zheng M J, Zhong S H, Liu Y. Steganographer detection via a similarity accumulation graph convolutional network. Neural Networks: The Official Journal of the International Neural Network Society, 2021, 136: 97?111 doi: 10.1016/j.neunet.2020.12.026 [10] Ning X, Tian W J, Yu Z Y, Li W J, Bai X, Wang Y B. HCFNN: High-order coverage function neural network for image classifican. Pattern Recognition, 2022, 131: Article No. 108873 doi: 10.1016/j.patcog.2022.108873 [11] Defferrard M, Bresson X, Vandergheynst P. Convolutional neural networks on graphs with fast localized spectral filtering. In: Proceedings of the 30th International Conference on Neural Information Processing Systems (NIPS). New York, USA: Curran Associates Inc., 2016. 3844?3852 [12] Kipf T N, Welling M. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv: 1609.02907, 2016. [13] Hamilton W L, Ying R, Leskovec J. Inductive representation learning on large graphs. In: Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS). New York, USA: Curran Associates Inc., 2017. 1025?1035 [14] Ying R, You J X, Morris C, Ren X, Hamilton W L, Leskovec J. Hierarchical graph representation learning with differentiable pooling. In: Proceedings of the 32nd International Conference on Neural Information Processing Systems (NIPS). New York, USA: Curran Associates Inc., 2018. 4805?4815 [15] Veli?kovi? P, Cucurull G, Casanova A, Romero A, Liò P, Bengio Y. Graph attention networks. In: Proceedings of the 6th International Conference on Learning Representation. Vancouver, Canada: ICLR Press, 2018. 1?12 [16] Gao W, Wan F, Yue J, Xu S C, Ye Q X. Discrepant multiple instance learning for weakly supervised object detection. Pattern Recognition: The Journal of the Pattern Recognition Society, 2022, 122: Article No. 108233 doi: 10.1016/j.patcog.2021.108233 [17] Tang X C, Liu M Z, Zhong H, Ju Y Z, Li W L, Xu Q. MILL: Channel attention-based deep multiple instance learning for landslide recognition. ACM Transactions on Multimedia Computing, Communications, and Applications, 2021, 17(2s): 1?11 [18] Yuan M, Xu Y T, Feng R X, Liu Z M. Instance elimination strategy for non-convex multiple-instance learning using sparse positive bags. Neural Networks, 2021, 142: 509?521 doi: 10.1016/j.neunet.2021.07.009 [19] Su Z Y, Tavolara T E, Carreno-Galeano G, Lee S J, Gurcan M N, Niazi M K K. Attention2majority: Weak multiple instance learning for regenerative kidney grading on whole slide images. Medical Image Analysis, 2022, 79: Article No. 102462 doi: 10.1016/j.media.2022.102462 [20] Bas P, Filler T, Pevny T. “Break our steganographic system”: The ins and outs of organizing BOSS. In: Proceedings of the International Workshop on Information Hiding. Prague, Czech Republic: Springer, 2011. 59?70 [21] Holub V, Fridrich J. Low-complexity features for JPEG steganalysis using undecimated DCT. IEEE Transactions on Information Forensics and Security, 2015, 10(2): 219?228 doi: 10.1109/TIFS.2014.2364918 [22] Fridrich J, Kodovsky J. Rich models for steganalysis of digital images. IEEE Transactions on Information Forensics and Security, 2012, 7(3): 868?882 doi: 10.1109/TIFS.2012.2190402 [23] Shi H C, Dong J, Wang W, Qian Y L, Zhang X Y. SSGAN: Secure steganography based on generative adversarial networks. In: Proceedings of the Pacific Rim Conference on Multimedia. Harbin, China: Springer, 2017. 534?544 [24] Pevny T, Somol P. Using neural network formalism to solve multiple-instance problems. arXiv preprint arXiv: 1609.07257, 2016. [25] Thekumparampil K K, Wang C, Oh S, Li L J. Attention-based graph neural network for semi-supervised learning. arXiv preprint arXiv: 1803.03735, 2018. [26] Xu G S, Wu H Z, Shi Y Q. Structural design of convolutional neural networks for steganalysis. IEEE Signal Processing Letters, 2016, 23(5): 708?712 doi: 10.1109/LSP.2016.2548421 [27] Holub V, Fridrich J, Denemark T. Universal distortion function for steganography in an arbitrary domain. EURASIP Journal on Information Security, DOI: 10.1186/1687-417X-2014-1 [28] Filler T, Fridrich J. Gibbs construction in steganography. IEEE Transactions on Information Forensics and Security, 2010, 5(4): 705?720 doi: 10.1109/TIFS.2010.2077629 [29] Holub V, Fridrich J. Designing steganographic distortion using directional filters. In: Proceedings of the IEEE International Workshop on Information Forensics and Security (WIFS). Piscataway, USA: IEEE, 2012. 234?239 [30] Sedighi V, Cogranne R, Fridrich J. Content-adaptive steganography by minimizing statistical detectability. IEEE Transactions on Information Forensics and Security, 2016, 11(2): 221?234 doi: 10.1109/TIFS.2015.2486744 [31] Fridrich J, Pevny T, Kodovsky J. Statistically undetectable JPEG steganography: Dead ends challenges, and opportunities. In: Proceedings of the 9th Workshop on Multimedia and Security. Dallas, USA: ACM, 2007. 3?14 [32] Guo L J, Ni J Q, Su W K, Tang C P, Shi Y Q. Using statistical image model for JPEG steganography: Uniform embedding revisited. IEEE Transactions on Information Forensics and Security, 2015, 10(12): 2669?2680 doi: 10.1109/TIFS.2015.2473815 [33] Qian Y L, Dong J, Wang W, Tan T N. Deep learning for steganalysis via convolutional neural networks. In: Proceedings of the SPIE 9409, Media Watermarking, Security, and Forensics. Bellingham, USA: SPIE, 2015. 1?10 [34] 衛星, 李佳, 孫曉, 劉邵凡, 陸陽. 基于混合生成對抗網絡的多視角圖像生成算法. 自動化學報, 2021, 47(11): 2623?2636Wei Xing, Li Jia, Sun Xiao, Liu Shao-Fan, Lu Yang. Cross-view image generation via mixture generative adversarial network. Acta Automatica Sinica, 2021, 47(11): 2623?2636 [35] 胡銘菲, 左信, 劉建偉. 深度生成模型綜述. 自動化學報, 2022, 48(1): 40?74 doi: 10.16383/j.aas.c190866Hu Ming-Fei, Zuo Xin, Liu Jian-Wei. Survey on deep generative model. Acta Automatica Sinica, 2022, 48(1): 40?74 doi: 10.16383/j.aas.c190866 [36] 董胤蓬, 蘇航, 朱軍. 面向對抗樣本的深度神經網絡可解釋性分析. 自動化學報, 2022, 48(1): 75?86Dong Yin-Peng, Su Hang, Zhu Jun. Interpretability analysis of deep neural networks with adversarial examples. Acta Automatica Sinica, 2022, 48(1): 75?86 [37] 余正飛, 閆巧, 周鋆. 面向網絡空間防御的對抗機器學習研究綜述. 自動化學報, 2022, 48(7): 1625?1649Yu Zheng-Fei, Yan Qiao, Zhou Yun. A survey on adversarial machine learning for cyberspace defense. Acta Automatica Sinica, 2022, 48(7): 1625?1649 [38] 趙博宇, 張長青, 陳蕾, 劉新旺, 李澤超, 胡清華. 生成式不完整多視圖數據聚類. 自動化學報, 2021, 47(8): 1867?1875 doi: 10.16383/j.aas.c200121Zhao Bo-Yu, Zhang Chang-Qing, Chen Lei, Liu Xin-Wang, Li Ze-Chao, Hu Qing-Hua. Generative model for partial multi-view clustering. Acta Automatica Sinica, 2021, 47(8): 1867?1875 doi: 10.16383/j.aas.c200121 [39] 張博瑋, 鄭建飛, 胡昌華, 裴洪, 董青. 基于流模型的缺失數據生成方法在剩余壽命預測中的應用. 自動化學報, 2023, 49(1): 185?196Zhang Bo-Wei, Zheng Jian-Fei, Hu Chang-Hua, Pei Hong, Dong Qing. Missing data generation method based on flow model and its application in remaining life prediction. Acta Automatica Sinica, 2023, 49(1): 185?196 [40] Wei G, Guo J, Ke Y Z, Wang K, Yang S, Sheng N. A three-stage GAN model based on edge and color prediction for image outpainting. Expert Systems With Applications, 2023, 214: Article No. 119136 doi: 10.1016/j.eswa.2022.119136 [41] Wang Y F, Dong X S, Wang L X, Chen W D, Zhang X J. Optimizing small-sample disk fault detection based on LSTM-GAN model. ACM Transactions on Architecture and Code Optimization, 2022, 19(1): 1?24 [42] Boroumand M, Chen M, Fridrich J. Deep residual network for steganalysis of digital images. IEEE Transactions on Information Forensics and Security, 2019, 14(5): 1181?1193 doi: 10.1109/TIFS.2018.2871749