Weak supervision JPEG image steganoer identification method and system based on graph kernel

Through a weak supervision method based on the graph kernel, suitable image features are extracted and user graph structure is constructed, combined with convolutional neural network, the oversmooth problem in the graph convolution process is solved, the accuracy of JPEG image steganography recognition is improved, and the realization of JPEG image steganography is realized. Effective identification of steganography.

CN119027725AActive Publication Date: 2024-11-26Chinese People's Liberation Army Cyberspace Force Information Engineering University
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411049558.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-26
Estimated Expiration
2044-08-01

AI Technical Summary

Technical Problem

The existing steganography recognition method based on graph neural network is prone to excessive smoothing problems during the construction and training of graph structures, which leads to indistinguishable characterizations. The image feature performance restricts the effectiveness of the recognition method, making it difficult to effectively identify JPEG. Steganizer in the image.

Method used

Using a weakly supervised JPEG image steganography recognition method based on graph kernels, the user graph structure is constructed by designing an image feature extraction strategy suitable for graph structures, and using graph kernels to map graph structure data to low-dimensional spaces, combined with convolutional nerves Online learning can capture the classification characteristics of user differences and thus identify steganography.

Benefits of technology

It effectively avoids the problem of oversmoothing in the graph convolution process, retains structural information and global attributes, improves the accuracy and effectiveness of steganizer recognition, and can more accurately identify steganizers in JPEG images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027725A_ABST
    Figure CN119027725A_ABST
Patent Text Reader

Abstract

The invention discloses a weak supervision JPEG (Joint Photographic Experts Group) image steganoer identification method and system based on a graph kernel, and the method comprises the steps: firstly analyzing a plurality of JPEG domain steganography detection feature methods, and selecting and extracting image features suitable for image structure steganoer identification; constructing the user into a graph structure by utilizing the features and combining structure information between the images; and finally, designing a user graph data standardization representation method based on a graph kernel, and learning user difference representation through the trained convolutional neural network. Experimental results show that the method provided by the invention has high recognition accuracy under various steganography embedding ratios for steganographers that embed secret information in an image by using various mainstream JPEG steganography methods such as nsF5, UED, J-UNIWARD and the like; when the DCTR3 feature of the image is extracted and a user graph structure is constructed based on the feature, the average recognition accuracy can be improved by nearly 20%, and therefore the effectiveness of the method is verified.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of network security technology, and in particular to a method and system for identifying JPEG image steganographers based on weak supervision of image kernels. Background Art

[0002] Steganography is a technology that embeds secret information into digital carriers such as images, audio, and video to achieve covert communication. The development of steganography plays an important role in ensuring the communication security of special departments and key groups of the country. However, while protecting the security of communications, if steganography is used by some malicious organizations to implement illegal covert communications, it will pose a serious threat to the security of cyberspace. How to effectively identify and monitor the covert communication behavior of criminals has become an urgent need for security departments in various countries.

[0003] Steganographer identification is an effective steganalysis technology, which aims to identify potential image steganographers on the network and discover illegal covert communication behaviors and subjects. Unlike steganalysis methods that usually focus on whether a single image is a secret image, steganographer identification is based on a batch of images as a whole to determine whether they contain secret information. This problem is considered to be one of the problems of "moving steganography and steganalysis from the laboratory to the real world". Compared with the single-image steganalysis method, the steganographer identification method needs to deal with the interference caused by the mismatch between steganography and the embedding rate of steganalysis on the one hand; on the other hand, steganographer identification needs to deal with the interference of secret images mixed in batches of carrier images on the recognition accuracy. All of the above put forward higher requirements for steganographer identification technology, but because of this, the development of steganographer identification technology is often more in line with the needs of practical applications.

[0004] In order to cope with the challenges brought by the algorithm mismatch problem such as steganography and steganalysis embedding rate, Ker et al. first proposed a steganalysis identification paradigm based on hierarchical clustering in 2011 [Ker AD, and T. Pevny. A new paradigm for steganalysis via clustering [C] / / Proc of Media Watermarking, Security, and Forensics, Manchester: 2011, 312–324]. The clustering method does not need to know the steganalysis algorithm and embedding rate used by the steganalyst, and does not require training, so it has better robustness. However, the premise of clustering is to assume that the steganalyst must exist. In practice, the steganalyst is not sure whether there is a steganalyst in the data to be tested. In this case, Ker et al. proposed a steganographer identification method based on local anomaly factors [Ker AD, Pevny T. The steganographer is the outlier: realistic large-scale steganalysis [J]. IEEE Transactions on Information Forensics and Security, 2014, 9 (9): 1424–1435]. This method uses a certain outlier to measure the degree of abnormality of the user. After that, steganographer identification methods based on different image features were proposed one after another. In addition to the differences in the identifiers, the performance of these methods mainly depends on the extracted image features. However, it is usually the case that in order to avoid being discovered by the steganalysis party, only a small part of the steganographer's image is often a secret image, that is, some secret images are mixed into the batch carrier images to interfere with the steganographer's identifier. At this time, the distribution of the steganographer's image features will be very similar to that of ordinary users, and it will become extremely difficult to identify the steganographer using only image features. In view of the series of important breakthroughs made by deep learning in graph structure learning, in 2020, Zhang et al. [Zhang Zhi, Zheng Mingjie, Zhong Shenhua, et al. Steganographer detection via enhancement-aware graph convolutional network [C] / / Proc of IEEE International Conference on Multimedia and Expo, Piscataway, NJ: IEEE, 2020: 1-6] proposed a steganographer identification method based on graph convolutional neural network. Specifically, after extracting image features, the features are further used to construct the user into a graph structure, and the steganographer is identified through the trained multi-layer graph convolutional network.

[0005] However, the image features need to be used in both the construction of the graph structure and the graph convolution training process, which means that the effectiveness of steganographer identification based on graph methods is largely restricted by the performance of the extracted image features. In addition, in order to learn classifiable user graph embeddings from the graph structure, graph convolutional neural networks often need to stack multiple layers of feature information, and the setting of this number of layers is usually difficult to grasp. If the number of layers is set too large, over-smoothing problems will occur, making the embeddings of deep graph nodes very similar. The stacking of these similar node embeddings will lead to a decrease in graph classification performance. Summary of the invention

[0006] The existing image steganographer identification method based on graph neural network captures this difference by training multi-layer graph convolution. On the one hand, due to the over-smoothing problem in graph neural network, it is very easy to make the learned representation indistinguishable, resulting in a decrease in recognition performance. On the other hand, the image feature performance largely restricts the effectiveness of the recognition method. In view of the above problems, the present invention proposes a weakly supervised JPEG image steganographer identification method and system based on graph kernel. Firstly, an image feature extraction strategy suitable for graph structure steganographer identification is designed. Under the guidance of this strategy, the adapted image features are extracted and the user graph structure is constructed; then, the graph kernel that can retain the structural information and global attributes to the greatest extent is studied, and a low-dimensional standardized representation of the user graph structure based on the graph kernel is obtained; finally, the user representation based on the graph kernel is integrated with the constructed convolutional neural network to learn the classification features that can capture user differences, thereby identifying the steganographer.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] On one hand, the present invention proposes a weakly supervised JPEG image steganographer identification method based on image kernel, comprising:

[0009] Step 1: Analyze the existing JPEG domain steganalysis detection features, select and extract JPEG image features suitable for graph structure steganalyzer identification;

[0010] Step 2: Using the extracted JPEG image features and combining the structural information between JPEG images, construct the user into a graph structure;

[0011] Step 3: Use the graph kernel to map the graph structure data corresponding to the user into a low-dimensional space to obtain the low-dimensional embedded data of the graph;

[0012] Step 4: Input the low-dimensional embedding data of the graph into the constructed convolutional neural network to obtain the final graph feature embedding, and perform classification prediction of the graph based on the final graph feature embedding to identify the JPEG image steganophobe.

[0013] Furthermore, the step 1 comprises:

[0014] This paper simulates users using steganography to embed secret information in JPEG images, and uses the maximum mean difference between users to perform statistical analysis on the JPEG image steganalysis detection features.

[0015] Furthermore, the maximum mean difference between users is calculated as follows:

[0016]

[0017] Where MMD(X,Y) represents the maximum mean difference between user X and user Y, x i and i are image samples of X and Y respectively, n is the number of samples, and k(x,y) is the kernel function.

[0018] Furthermore, the step 2 comprises:

[0019] JPEG images are used as nodes, similarities between JPEG images are used as edges, and users are constructed into a graph structure.

[0020] Furthermore, in step 2, the similarity between JPEG images is calculated using a Gaussian kernel function:

[0021]

[0022] Where s(f i,u ,f i,v ) represents user x i The similarity between image nodes u and v in the corresponding graph structure, f i,u ,f i,v are the eigenvectors of u and v respectively, σ represents the range control coefficient of the Gaussian kernel function, ||·|| 2 represents the L2 norm.

[0023] Further, in step 3, based on Method to obtain The rows of Q are low-dimensional embeddings of the graph, N is the number of users, and p is the vector dimension. For the graph G′, the mapping is calculated as follows

[0024]

[0025] Where z′ represents the low-dimensional embedded data after the mapping of graph G′, yes The pseudo-reversal, is the kernel value between G′ and other graphs in the dataset.

[0026] Furthermore, the classification prediction of the graph based on the final graph feature embedding includes:

[0027] The maximum pooling operation is applied on the graph feature embedding to retain only the maximum value in the graph feature embedding, which is used as the signal associated with the filter. Finally, two fully connected layers are connected and the Softmax function is used to complete the classification prediction of the graph.

[0028] On the other hand, the present invention proposes a weakly supervised JPEG image steganographer identification system based on image kernel, comprising:

[0029] An image feature extraction unit, used for analyzing existing JPEG domain steganalysis detection features, selecting and extracting JPEG image features suitable for graph structure steganalyzer identification;

[0030] A user graph construction unit, configured to construct the user into a graph structure by utilizing the extracted JPEG image features and combining the structural information between JPEG images;

[0031] A user graph structure mapping unit, used to map the graph structure data corresponding to the user into a low-dimensional space using the graph kernel to obtain low-dimensional embedding data of the graph;

[0032] The user classification and identification unit is used to input the low-dimensional embedding data of the graph into the constructed convolutional neural network to obtain the final graph feature embedding, and perform classification prediction of the graph based on the final graph feature embedding, thereby identifying the JPEG image stegangrapher.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention first designs an image feature extraction strategy suitable for graph structure steganography identification. Under the guidance of this strategy, the adapted image features are extracted and the user graph structure is constructed. Then, the graph kernel that can retain the structural information and global attributes to the greatest extent is studied, and a low-dimensional standardized representation of the user graph structure based on the graph kernel is obtained. Finally, the user representation based on the graph kernel is integrated with the constructed convolutional neural network to learn the classification features that can capture user differences, thereby identifying JPEG image steganophobes.

[0035] The present invention analyzes and selects image features suitable for graph structure construction in JPEG image steganographer identification. Various JPEG image steganalysis detection feature design methods are deeply analyzed to select image features that are more suitable for graph structure-based steganographer identification tasks.

[0036] In view of the advantage of graph kernel in preserving structural information in low-dimensional mapping of graph structure, this paper introduces it into steganographer identification to characterize users, and proposes a weakly supervised JPEG image steganographer identification method based on graph kernel, avoiding the problem of user information loss caused by over-smoothing in the graph convolution process in existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1This is a flow chart of a method for identifying steganophobes of JPEG images based on weak supervision of image kernels according to an embodiment of the present invention;

[0038] Figure 2 A statistical diagram of the difference measurement between users provided by an embodiment of the present invention;

[0039] Figure 3 Recognition results represented by different image features provided by the embodiments of the present invention;

[0040] Figure 4 The present invention is a schematic diagram of the architecture of a weakly supervised JPEG image steganographer identification system based on image kernel according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0042] The present invention constructs each user as a graph structure (Graph), and the steganographer identification problem is transformed into a graph classification task. Graph includes nodes and edges. In the constructed user graph structure, the nodes correspond to the user's images, and the edges describe the similarity relationship between images; at this time, the embedding of the nodes is represented by the features of the image, and the edges are represented by the distance similarity between the features. From the bottom layer, the construction of the graph depends on the representation of the image features, which means that in addition to the design of the graph classification model, the image feature performance largely restricts the effectiveness and quality of the steganographer identification based on the graph method. At this time, a suitable image feature is particularly important for the model. Based on this, the present invention proposes an image feature selection strategy suitable for steganographer identification based on an in-depth analysis of a variety of JPEG image steganalysis features, selects a suitable image feature representation based on the strategy, and finally proposes a weakly supervised JPEG image steganographer identification method based on the graph kernel. Figure 1 As shown, the method includes:

[0043] Step 1: Analyze the existing JPEG domain steganalysis detection features, select and extract JPEG image features suitable for graph structure steganalyzer identification;

[0044] Step 2: Using the extracted JPEG image features and combining the structural information between JPEG images, construct the user into a graph structure;

[0045] Step 3: Use the graph kernel to map the graph structure data corresponding to the user into a low-dimensional space to obtain the low-dimensional embedded data of the graph;

[0046] Step 4: Input the low-dimensional embedding data of the graph into the constructed convolutional neural network to obtain the final graph feature embedding, and perform classification prediction of the graph based on the final graph feature embedding to identify the JPEG image steganophobe.

[0047] The method specifically includes:

[0048] 1 Image feature analysis and extraction for steganographer identification

[0049] Steganographic identification uses image features to measure the differences between users, usually using the maximum mean discrepancy (MMD) as a distance metric. Specifically, the MMD distance between users X and Y is

[0050]

[0051] Among them, x i and i are image samples of X and Y respectively, and n is the number of samples. The kernel function k(x,y) maps the original variables to a high-dimensional space. Commonly used kernel functions include linear kernels

[0052] k(x,y)=x T y (2)

[0053] and Gaussian kernel

[0054] k(x,y)=exp(-γ||xy|| 2 ) (3)

[0055] Here, γ represents the width of the kernel.

[0056] Based on this, in order to seek image features suitable for stegographer identification, the present invention analyzes the existing typical JPEG domain steganalysis detection features. Specifically, firstly, it simulates a user using steganography to embed secret information in his own image, and the user is the stegographer. Two steganography techniques were tested experimentally, including image non-adaptive nsF5 steganography and adaptive J-UNIWARD (JPEG UNIversalWAveletRelativeDistortion) steganography. Then, the JPEG image steganalysis detection features were statistically analyzed using formula (1). It is worth noting that the purpose of the feature analysis here is to seek image features that can reflect user differences, so all images of the stegographer are set to be secret images.

[0057] Unlike other image classification tasks, the design of image steganalysis features focuses on understanding the smaller steganalysis noise or feature changes in the image, while suppressing the influence of the image content. JPEG (joint photographic experts group) is a commonly used image format on the Internet. It has the characteristics of sufficient redundant space and widespread application, and is a good carrier. The PEV-274 feature proposed by Pevny et al. [Pevny T, Fridrich J. Multiclass detector of current steganographic methods for JPEG format [J]. IEEE Transactions on Information Forensics and Security, 2008, 3 (4): 635-650] can reflect the impact of various types of early non-adaptive JPEG image steganographic embedding, while greatly reducing the interference of content on steganalysis, and achieving good detection results. The feature has a total of 274 dimensions. The CC-PEV feature [Kodovsky J, Fridrich J. Calibration revisited [C] / / Proc of the ACM Workshop on Multimedia and security, New York: ACM: 2009: 63-74] is an improved version of PEV-274 that incorporates Cartesian calibration. Its dimension is twice that of PEV-274, 548 dimensions. With the emergence of STCs (syndrome-trellis codes) coding technology, more advanced adaptive JPEG steganography was born. Adaptive steganalysis embedding is mainly concentrated in complex areas that are difficult to model and are more difficult to detect. For adaptive JPEG steganography, the DCTR (DCT residual) feature, as a breakthrough work, has opened up a new framework for JPEG phase-aware features. However, the DCTR feature dimension is also much higher than PEV-274 and CC-PEV, reaching an astonishing 8,000 dimensions. The DCTR that was born in recent years M(M=3,4,8) series features [Xia Chao, Guan Qingxiao, Zhao Xianfeng, et al. Improved JPEG phase-aware steganalysis features using multiple filter sizes and difference images [J]. IEEE Transactions on Circuits and Systems for Video Technology, 2020, 30(11): 4100-4113] also borrow the design idea of ​​DCTR feature, but this feature obtains residual by setting filters of different sizes (M represents the filter size), which can reduce the feature dimension while capturing the local texture changes before and after adaptive JPEG steganalysis embedding. Among them, when M=3, DCTR 3 The feature is 390-dimensional; when M=4, DCTR 4 The feature is 1050-dimensional; when M=8, it is 8000-dimensional DCTR feature. However, due to the use of a feature symmetry method different from DCTR, DCTR 8 The feature dimension is only 4100, which is about 1 / 2 of the original DCTR feature dimension.

[0058] According to the different image feature design methods, three low-dimensional steganalysis detection features PEV, CC-PEV and DCTR are analyzed. 3 The specific approach is: based on the BOSSBase-1.01 image dataset commonly used in the field of steganalysis and steganography, 10,000 JPEG images with a quality factor of QF (quality factor) of 75 are generated as carrier images, and the corresponding steganographic images are generated using the nsF5 and J-UNIWARD steganographic algorithms with an embedding rate of 0.5bpnzAC (bits per nonzero AC DCT coefficient). A certain number of images are randomly selected from the carrier image set as normal user images, and the same number of images are randomly selected from the steganographic image set to form the stegographer's images. The PEV, CC-PEV and DCTR of all images are extracted respectively. 3 Steganalysis detection features, using formula (1) to measure the advantages and disadvantages of these three features in measuring the differences between users. The experimental results are as follows Figure 2 As shown, Figure 2 (a) is the measurement result of the steganographer using nsF5 steganography. Figure 2(b) is the measurement result of the steganography by the steganographic user using J-UNIWARD steganography. The horizontal axis represents the number of images of each user, and the vertical axis represents the MMD distance measurement result between users. The present invention tests the number of images of 10 specifications for each user (from 10 to 100 images), and uses MMD to repeat the calculation 100 times and take the average value. Figure 2 It can be seen that when measuring the difference between users, using DCTR is more effective than PEV and CC-PEV features in both steganography techniques. 3 The differences between users measured by the feature are more significant. After analysis and selection of applicable image features, the features can be used for subsequent recognition tasks.

[0059] 2. Structure-aware User Graph Construction

[0060] The images are used as nodes and the similarities between images are used as edges to construct a structure-aware user graph structure. The similarity between images is calculated using the Gaussian kernel function. Specifically, given a user x containing multiple images, i , x i The similarity between any two image nodes u and v in is expressed as

[0061]

[0062] Among them, f i,u ,f i,v are the eigenvectors of u and v respectively, σ controls the range of the Gaussian kernel function, ||·|| 2 Represents the L2 norm, which is the Euclidean distance in the usual sense, that is

[0063]

[0064] Among them, d represents the dimension of node features.

[0065] Userx i All image node features constitute the feature matrix of the user n is user x i The number of images, d is the image feature dimension. Using formula (4), metric learning is performed on this feature matrix to obtain image similarity As x i The adjacency matrix of Right now

[0066]

[0067] Among them, s(·) is the similarity function, f i,u ,f i,v are the feature vectors of any two nodes u and v, respectively. is an indicator function. When the condition is met, its value is 1, otherwise it is 0. The threshold τ∈[0,1] is a set hyperparameter. The average distance between all nodes in the graph can also be taken as the threshold. i and the adjacency matrix A i After that, user x i It is represented as a graph structure. At this point, the user classification and recognition problem is transformed into a graph similarity problem.

[0068] 3. Steganographic identification based on graph core

[0069] The present invention utilizes a graph kernel to embed the user graph dataset into a low-dimensional space, so that in this low-dimensional space, the structural information and properties of the graph can still be retained to the maximum extent.

[0070] Let G = {G i :i∈{1,2,...,N}} is the user graph set and N is its cardinality. is the kernel matrix constructed by G using the graph kernel K. The algorithm complexity of constructing the complete graph kernel matrix K and factorizing it to obtain the low-dimensional embedding of the graph is O(N 3 ), which is computationally expensive. The method is considered to be the most representative kernel matrix low-rank approximation method. The method obtained The kernel matrix can be obtained Its complexity is O(N 3 ) is reduced to O(p 2 N), where p<<N. By using only a small subset of the p columns (or rows) of the kernel matrix, the computational cost of the kernel matrix is ​​reduced. At this time, the rows of Q are the low-dimensional embeddings of the graph, which are used as the standardized user graph vectors for the input of training subsequent convolutional neural networks, with a vector dimension of p.

[0071] When testing a new user graph, it is mapped to the feature space of the graph in the training set. Specifically, given a new user graph G′, its mapping is calculated as follows

[0072]

[0073] in, yes The pseudo-reversal, is the kernel value between G′ and other graphs in the training set. The dimension of the mapped vector is consistent with the dimension of the normalized user graph vector in the training set. After performing the above operations on all user subgraphs, these graphs are finally embedded into the input constructed convolutional neural network for downstream user classification tasks.

[0074] Assume z i For Figure Gi The standardized embedding of the feature c after the convolution filter operation i for

[0075] c i =σ(w i z i +b i ) (8)

[0076] Among them, σ is the activation function, ReLU is used here, w i is the filter learned by the network. Any filter learned by the network All of them belong to the dot product of the graph kernel K to the regenerated Hilbert space H, that is,

[0077] <z i ,z j > = k(G i ,G j )=<φ(G i ),φ(G j )> H (9)

[0078] Among them, z i ,z j For Figure G i and G j The standardized feature vector of .

[0079] After performing multi-layer convolution on the standardized vectorization of all user graphs, the feature embedding of the following graph G is obtained:

[0080] c=[c 1 ,c 2 ,...,c N ] T (10)

[0081] In the final classification module of the network, a max-pooling operation is applied to the user graph embedding. The max-pooling operation only retains the maximum value in the feature embedding (mapping) c, which is used as a signal related to the filter w. An intuitive understanding is that the max-pooling operation is beneficial to capturing the feature representation of the class to which the graph belongs. Finally, two fully connected layers FC (fully connected) are connected and the Softmax function is used to complete the user graph classification prediction and identify the JPEG image stegosaurus, that is,

[0082]

[0083] 4. Experiment

[0084] 4.1 Experimental Setup

[0085] The experiment was conducted on the BOSSBase-1.01 image database, a commonly used benchmark dataset in the field of steganalysis. The BOSSBase-1.01 database contains 10,000 512×512 pixel grayscale uncompressed natural images. Since our research object is JPEG images, we first use the "imwrite" function of MATLAB to compress the images into JPEG images with a quality factor QF=75 as experimental carrier images. Due to the wide application of JPEG domain steganographic algorithms nsF5, UED and J-UNIWARD in the field of information hiding, the present invention uses these three steganography techniques to construct secret images with embedding rates of 0.2bpnzAC, 0.3bpnzAC, 0.4bpnzAC and 0.5bpnzAC on the carrier images respectively.

[0086] In addition to the steganographic carrier, steganographic algorithm and embedding rate, the research object of this invention is the user who embeds secret information on multiple images, which also involves a batch embedding strategy problem, that is, what strategy should the steganologist choose to embed the secret information of length Mbit into n available carriers (I 1 ,I 2 ,…,I n The experiment of the present invention adopts a uniform embedding strategy, that is, the steganographic writer embeds the secret information uniformly into the carriers available to him, without considering the embedding capacity of a single carrier.

[0087] 4.2 Comparative experimental results based on different image features

[0088] The performance of different image features in the model proposed in this section is mainly to evaluate the image feature performance based on the same model when users share images of different numbers and specifications, including PEV, CC-PEV, DCTR and DCTR 3 Four image steganalysis features.

[0089] Specifically, firstly, two steganography techniques (nsF5 and J-UNIWARD) with a set embedding rate (here we choose a high embedding rate of 0.5bpnzAC) are used to generate corresponding secret images using the carrier image, where the carrier image is used as the image dataset of normal users and the secret image is used as the image dataset of the steganist. Then, the carrier and secret images are used to simulate normal network users and steganists (network image secret communication subjects). Each user is set to share a certain number of images (including six specifications of 5, 10, 20, 30, 40, and 50 images), and 200 normal users and 200 steganists are constructed for each specification. For example, the first specification is that each user shares 5 images, with a total of 400 users (200 normal users and 200 steganists), and 2,000 images are needed (including 1,000 carrier images and 1,000 secret images), and the rest of the rules are set in the same way. It should be noted that the normal user images are randomly selected from the carrier image set, and the steganographic images are randomly selected from the steganographic images. In the experiment, 90% of the constructed user data is used as the training set for model testing, and the remaining 10% is used as the test set. The training set uses 10 times of 10-fold cross validation, that is, the training data set is divided into 10 parts, and 1 part is taken as the validation set each time, and the remaining 9 parts are used as the training set, and then the results of these 10 times are averaged. The experimental results are shown in Figure 3 shown.

[0090] Figure 3 The test accuracy of different image feature extraction methods in the constructed model is intuitively demonstrated. Figure 3 (a) is the identification result of the steganographer using the nsF5 steganography model. Figure 3 (b) is the identification result of the stegostudent using the J-UNIWARD steganography model. Figure 3 In the figure, the horizontal axis represents the number of images shared by users, which are 5, 10, 20, 30, 40, and 50 images respectively, and the vertical axis represents the recognition accuracy of the model. Figure 2 It can be clearly seen that different image features perform well in the steganographer recognition model. When users use different steganography techniques to embed secret information in batches of images, DCTR 3The performance of these features is significantly better than other image features, especially when the user adopts adaptive steganography, the advantage of this feature is more obvious. We analyzed that this is mainly because, on the one hand, the ability of PEV and CC-PEV features to capture adaptive steganographic embedding is relatively limited, resulting in unsatisfactory recognition performance of such features when users use adaptive steganography; on the other hand, although the DCTR feature can capture adaptive steganographic embedding, its dimension is too high. The high feature dimension causes the distance measurement to fail, that is, it presents the well-known dimensionality curse problem when measuring distance, making the performance of using this feature for stegostalker identification even worse than that of PEV and CC-PEV features. 3 The feature has the advantages of both detection effectiveness and low dimensionality, which makes it perform better than other methods in the model, and also provides reference for further research on the problem of steganographer identification. Therefore, in the comparative experiments of our model and other models, DCTR is selected unless otherwise specified and when necessary. 3 feature.

[0091] In summary, this paper proposes a weakly supervised JPEG image steganographer identification method based on graph kernel. This method analyzes the image feature extraction strategy suitable for graph structure steganographer identification, and uses the DCTR of the image under the guidance of this strategy. 3 The user graph structure is constructed based on the features; then, the user graph embedding obtained based on the graph kernel is convolved with the one-dimensional convolutional neural network filter and passed to the neural network for training to complete the graph classification, thereby identifying the steganographic operator. A series of experiments were conducted on the commonly used BOSSbase-1.01 image database. The experimental results show that the image feature representation method proposed in this invention can effectively identify the steganographic operator who uses JPEG images for covert communication.

[0092] Based on the above embodiments, Figure 4 As shown, the present invention also proposes a weakly supervised JPEG image steganographer identification system based on image kernel, comprising:

[0093] An image feature extraction unit, used for analyzing existing JPEG domain steganalysis detection features, selecting and extracting JPEG image features suitable for graph structure steganalyzer identification;

[0094] A user graph construction unit, configured to construct the user into a graph structure by utilizing the extracted JPEG image features and combining the structural information between JPEG images;

[0095] A user graph structure mapping unit, used to map the graph structure data corresponding to the user into a low-dimensional space using the graph kernel to obtain low-dimensional embedding data of the graph;

[0096] The user classification and identification unit is used to input the low-dimensional embedding data of the graph into the constructed convolutional neural network to obtain the final graph feature embedding, and perform classification prediction of the graph based on the final graph feature embedding, thereby identifying the JPEG image stegangrapher.

[0097] In summary, the present invention analyzes and selects image features suitable for graph structure construction in JPEG image steganographer identification. In-depth analysis of various JPEG image steganalysis detection feature design methods is performed to select image features that are more suitable for graph structure-based steganographer identification tasks. In view of the advantage of graph kernels in preserving structural information in low-dimensional mapping of graph structures, the present invention introduces them into steganographer identification to characterize users, and proposes a weakly supervised JPEG image steganographer identification method based on graph kernels. Avoid the problem of user information loss caused by over-smoothing during graph convolution in existing methods.

[0098] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A weakly supervised JPEG image steganographer identification method based on image kernel, characterized in that: include: Step 1: Analyze the existing JPEG domain steganalysis detection features, select and extract JPEG image features suitable for graph structure steganalyzer identification; Step 2: Using the extracted JPEG image features and combining the structural information between JPEG images, construct the user into a graph structure; Step 3: Use the graph kernel to map the graph structure data corresponding to the user into a low-dimensional space to obtain the low-dimensional embedded data of the graph; Step 4: Input the low-dimensional embedding data of the graph into the constructed convolutional neural network to obtain the final graph feature embedding, and perform classification prediction of the graph based on the final graph feature embedding to identify the JPEG image steganophobe.

2. According to the weakly supervised JPEG image steganographer identification method based on image kernel in claim 1, it is characterized in that: The step 1 comprises: This paper simulates users using steganography to embed secret information in JPEG images, and uses the maximum mean difference between users to perform statistical analysis on the JPEG image steganalysis detection features.

3. According to the weakly supervised JPEG image steganographer identification method based on image kernel in claim 2, it is characterized in that: The maximum mean difference between users is calculated as follows: Where MMD(X,Y) represents the maximum mean difference between user X and user Y, x i and i are image samples of X and Y respectively, n is the number of samples, and k(x,y) is the kernel function.

4. According to the weakly supervised JPEG image steganographer identification method based on image kernel in claim 1, it is characterized in that: The step 2 comprises: JPEG images are used as nodes, similarities between JPEG images are used as edges, and users are constructed into a graph structure.

5. According to the weakly supervised JPEG image steganographer identification method based on image kernel according to claim 4, it is characterized in that: In step 2, the Gaussian kernel function is used to calculate the similarity between JPEG images: Where s(f i,u ,f i,v ) represents user x i The similarity between image nodes u and v in the corresponding graph structure, f i,u ,f i,v are the eigenvectors of u and v respectively, σ represents the range control coefficient of the Gaussian kernel function, ||·|| 2 represents the L2 norm.

6. The weakly supervised JPEG image steganographer identification method based on image kernel according to claim 1, characterized in that: In step 3, based on Method to obtain The rows of Q are low-dimensional embeddings of the graph, N is the number of users, and p is the vector dimension. For the graph G′, the mapping is calculated as follows Where z′ represents the low-dimensional embedded data after the mapping of graph G′, yes The pseudo-reversal, is the kernel value between G′ and other graphs in the dataset.

7. The weakly supervised JPEG image steganographer identification method based on image kernel according to claim 1, characterized in that: The classification prediction of the graph based on the final graph feature embedding includes: The maximum pooling operation is applied on the graph feature embedding to retain only the maximum value in the graph feature embedding, which is used as the signal associated with the filter. Finally, two fully connected layers are connected and the Softmax function is used to complete the classification prediction of the graph.

8. A weakly supervised JPEG image steganographer identification system based on image kernel, characterized in that: include: An image feature extraction unit, used for analyzing existing JPEG domain steganalysis detection features, selecting and extracting JPEG image features suitable for graph structure steganalyzer identification; A user graph construction unit, configured to construct the user into a graph structure by utilizing the extracted JPEG image features and combining the structural information between JPEG images; A user graph structure mapping unit, used to map the graph structure data corresponding to the user into a low-dimensional space using the graph kernel to obtain low-dimensional embedding data of the graph; The user classification and identification unit is used to input the low-dimensional embedding data of the graph into the constructed convolutional neural network to obtain the final graph feature embedding, and perform classification prediction of the graph based on the final graph feature embedding, thereby identifying the JPEG image stegangrapher.

Citation Information

Patent Citations

  • Image steganalysis method and system based on convolutional neural network

    CN108961137A

  • A convolutional neural network-based JPEG image steganography analysis method

    CN109934761A