CSM-oriented steganalysis network training method and device based on progressive intermediate domain

Abstract

This invention discloses a method and apparatus for training a steganalysis network based on a progressive intermediate domain (FDM) for a CSM. It proposes an intermediate domain generation module (IDGM) based on local feature hybridization (LFMP). By generating an intermediate domain, the source and target domains are effectively bridged, thereby enhancing the generalization capability of the steganalysis network and alleviating the degradation of the CSM's detection accuracy for the target domain. By sampling domain factors from the distribution Beta(α, 1) to control the bias of the intermediate domain, the generated intermediate domain gradually shifts from a distribution close to the source domain to one close to the target domain. This facilitates the gradual transfer of knowledge from the source to the target, further improving the steganalysis network's detection accuracy for the target domain.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a CSM-oriented steganalysis network training method and device based on a progressive intermediate domain. Background Art

[0002] Image steganography is a covert communication technique that embeds secret information into a carrier image without attracting the attention of anyone other than the sender and receiver. Image steganalysis, the inverse of image steganography, is a technique for detecting whether a carrier image has been steganographically encoded.

[0003] Previous steganalysis detectors aimed to analyze the statistical differences in handcrafted features between the carrier and cryptanalyst images through traditional signal processing and machine learning. However, with the increasing sophistication of steganographic techniques, manually defined features have become challenging, hindering the further development of traditional steganalysis algorithms. Recently, thanks to deep learning's ability to automatically extract complex statistical dependencies from images while simultaneously improving detection performance, a number of high-performance steganalysis detectors based on deep learning have been proposed.

[0004] Existing steganalysis detectors achieve high detection accuracy in laboratory scenarios. However, in the real world, the images in the training and test sets come from different sources and therefore have different distributions, which inevitably leads to a significant reduction in the detection accuracy of steganalysis detectors. This phenomenon is called source mismatch.

[0005] Although J-Net has been proposed to alleviate the CSM problem in deep steganalysis, J-Net achieves domain alignment by reducing the domain difference between the training set (source domain) and the test set (target domain), thereby solving the problem of decreased detection accuracy in deep steganalysis caused by CSM, J-Net ignores the extreme values and extremely different distributions of the source and target domains. This situation greatly hinders the direct transfer of knowledge from the source to the target, making the classifier trained on the source poorly adapted to the target domain. Therefore, the improvement in target domain detection accuracy brought by J-Net is limited. In addition, J-Net is only applicable to spatial domain steganalysis and cannot solve the CSM problem that also exists in JPEG domain steganalysis. Summary of the Invention

[0006] The purpose of the present invention is to provide a CSM-oriented steganalysis network training method and device based on a progressive intermediate domain.

[0007] The technical solution adopted in the present invention is:

[0008] A steganalysis network training method for CSM based on progressive intermediate domain is provided. The steganalysis network IDGM-Snet based on progressive intermediate domain includes a first convolution module, an intermediate domain generation module IDGM, a second convolution module and a classifier. The first convolution module, the second convolution module and the classifier constitute the backbone SRNet of IDGM-SNet. The first convolution module receives the source domain image and the target domain image and outputs the source hidden features and the target hidden features; the intermediate domain generation module IDGM is used to generate local mixing positions. The intermediate domain generation module IDGM performs local feature mixing LFMP on the source hidden features and the target hidden features output by the first convolution module, and uses the domain factor λ to control the mixing ratio of the local feature mixing LFMP, thereby generating intermediate domain features that gradually approach the target domain. The local feature mixing LFMP is used to locally mix the features of the source domain and the target domain; the second convolution module extracts features from the source domain, the target domain and the intermediate domain, and the gradually adapted classifier classifies the extracted features; the training method includes the following steps:

[0009] Step 1: Set the number N s The number of labeled source domain images is N t The unlabeled target domain image is input into the steganalysis network SRNet, and the source domain is output The intermediate layer features f i s (output of the ninth feature extraction layer), predicted label Output target domain The middle layer features (output of the ninth feature extraction layer), predicted label in, represents the i-th image in the source domain image set, represents the label of the i-th image in the source domain image set; represents the jth image in the target domain image set, 0 means the image is a carrier image, and 1 means the image is a secret image.

[0010] Step 2: Determine the intermediate layer features f corresponding to the source domain and the target domain i s and The local mixed position of , that is, generating a binary mask M of size w×v, the binary mask M includes multiple blocks of size c×c and all elements are 1; c represents the side length of the block with element 1;

[0011] Step 3: Expand the single-channel binary mask M of size w×v into an expanded mask M′ of size w×v×C, and compare the expanded mask M′ with f i s and The corresponding position elements of are multiplied to obtain the source patch features for the blending operation and target patch features Where w and v represent length and width respectively, and C is the number of channels; the specific formula is as follows:

[0012]

[0013] Among them, ⊙ represents the multiplication of the elements at corresponding positions in the matrix.

[0014] Step 4: Introduce the domain factor λ to adjust the source patch characteristics and target patch features The degree of influence on the intermediate domain features is used to obtain the domain factor mixed patch features Right now:

[0015]

[0016] in, In short, the closer λ is to 0, the Closer to the source domain features. On the contrary, λ closer to 1 means Closer to the target domain features. During training, the domain factor λ in step 4 is adjusted from the distribution Beta(α,1) and satisfies Where τ and T represent the current training iteration and the total training iterations respectively.

[0017] Step 5: For the zero-value region in M′ that excludes local mixing, select the retained source feature f i s The corresponding area of ​​f i s and Multiply the corresponding positions to obtain the source-retained features that retain some of the source features

[0018]

[0019] Where I is a matrix of all 1s,

[0020] Step 6: Source Preservation Features Combined domain factor hybrid patch features Calculate the intermediate domain feature f i m :

[0021]

[0022] in, Indicates that the source retains characteristics.

[0023] Step 7: According to the intermediate domain feature f im The specific steps are as follows:

[0024] Step 7-1, calculate the mixed patch features The label of M′ is denoted as a, and the proportion of the area with 1 elements in M′ to the whole is obtained. Tagged Specifically, based on formula (3), we get:

[0025]

[0026] Step 7-2, calculate source preservation characteristics Tags based on The complementary characteristics of M′ give The proportion of the area with the element 1 in the whole is 1-a, then the source retains the characteristics The label is

[0027] Step 7-3, calculate the intermediate domain feature f i m The derived tags are:

[0028]

[0029] Step 8: Calculate the overall loss function value based on the derived label The overall loss function value includes the cross entropy loss function value on the source domain and the cross entropy loss function value on the intermediate domain:

[0030]

[0031] Where β is a balance parameter, is the binary cross entropy loss function, which is as follows:

[0032]

[0033] in, and y represent the predicted label and true label of the sample, respectively.

[0034] Step 9, based on the overall loss function value Train the steganalysis network.

[0035] Furthermore, the steps for generating the binary mask M in step 2 are as follows:

[0036] Step 2-1, prepare a random binary matrix M0 using w×v binary numbers randomly drawn from the Bernoulli distribution B(1,q):

[0037]

[0038] in, represents an element of M0, k1 and k2 represent the row and column of the element respectively;

[0039] Step 2-2, each When it is equal to 1, a c×c square block centered at (k1, k2) is generated, and each element of the square block is set to 1, and finally a binary mask M is obtained.

[0040] Obviously, visiting each non-zero point of M0 in turn (i.e. ), the generated square blocks with element value 1 inevitably overlap with each other, which directly leads to the generation of irregular areas with element value 1.

[0041] This paper adopts the above technical solutions and proposes an IDGM based on local feature hybrid patchup (LFMP). By generating an intermediate domain, it effectively bridges the source and target domains, thereby enhancing the generalization ability of IDGM-Snet and alleviating the degradation of target domain detection accuracy caused by CSM. This paper controls the bias of the intermediate domain by sampling domain factors from the distribution Beta(α,1), so that the generated intermediate domain gradually transfers from a distribution close to the source domain to a distribution close to the target domain. This helps to gradually transfer knowledge from the source to the target, further improving the detection accuracy of IDGM-SNet for the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0043] Figure 1 Schematic diagram of the CSM-oriented steganalysis network structure based on the progressive intermediate domain of the present invention;

[0044] Figure 2 It is a visualization diagram of binary matrices M0 and M;

[0045] Figure 3 Schematic diagram of the local feature hybrid LFMP structure;

[0046] Figure 4 Schematic diagram of Beta(α,1) distribution. Implementation Method

[0047] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0048] like Figures 1 to 4As shown in one, the present invention discloses a CSM-oriented steganalysis network training method and device based on a progressive intermediate domain, and proposes a high-performance steganalysis detector IDGM-SNet that solves the carrier-source mismatch problem. The core of IDGM-SNet is a plug-and-play intermediate domain generation module (IDGM), which includes generating local mixing positions, local feature-level mixing correlation patches, and domain factors sampled from the distribution Beta(α,1). IDGM-SNet locally mixes the source domain and target domain features by sampling domain factors that gradually change from 0 to 1, generating an intermediate domain that gradually shifts from a distribution close to the source domain to a distribution close to the target domain, gradually bridging the extreme source domain and target domain, thereby improving the steganalysis accuracy on the target domain.

[0049] like Figure 1 As shown, the gray dots represent the area with a value of 1, and the white dots represent the area with a value of 0. The source domain features and target domain features in the feature map are marked with A and B respectively, and the new features obtained by locally mixing the source domain features and the target domain features are marked with D. IDGM-Snet mainly consists of four parts: the first convolution module, the intermediate domain generation module IDGM, the second convolution module and the classifier. Among them, the first convolution module, the second convolution module and the classifier constitute the backbone SRNet of IDGM-SNet. In the rest of this article, for simplicity, represents the first convolutional module before IDGM, and It is used to represent the second convolution module and the classifier. Local Feature Mixture LFMP is used to locally mix the features of the source domain and the target domain, which is a key part of IDGM. The goal of IDGM is to generate local mixed locations. The output source and target latent features undergo local feature blending (LFMP), and the domain factor λ is used to control the blending ratio of the local feature blending (LFMP), thereby generating intermediate domain features that gradually approach the target domain. Subsequently, the source, target, and intermediate domain features are combined in a second convolutional module for feature extraction. Finally, a gradually adapted classifier is used for classification. The specific steps are as follows.

[0050] (1) Set the quantity to N s The number of labeled source domain images is N t The unlabeled target domain image is input into the steganalysis network SRNet, and the source domain and target domain images are For example, the intermediate layer features f of the source domain and the target domain are output respectively i s , (output of the ninth feature extraction layer), predicted label in, 0 means the image is a carrier image, and 1 means the image is a secret image.

[0051] (2) Determine the source domain and target domain features f i s , The local mixing position. Essentially, determining the local mixing position is the process of generating a binary mask M of size w×v, which contains multiple blocks of size c×c, whose elements are all 1. The detailed generation process of M is mainly divided into the following two steps. Interestingly, f i s , All C channels of correspond to the same mask M, where f i s , The size of is w×v×C, so the size of M is w×v.

[0052] (2.1) Prepare a random binary matrix using w×v binary numbers randomly drawn from the Bernoulli distribution B(1,q) in, represents an element of M0, k1 and k2 represent the row and column of the element respectively; See Figure 2 Figure (a).

[0053] (2.2) When When , a c×c square block centered at (k1, k2) is generated, and each element is set to 1. Obviously, each non-zero point of M0 (i.e. ), the generated square blocks with element values of 1 inevitably overlap with each other, directly leading to the generation of irregular areas with element values of 1. Figure 2 Figure (b).

[0054] like Figure 2 Figure 2 shows the visualization of binary matrices M0 and M, where the white-marked points represent their values 0 and the black-marked points represent their values 1.

[0055] (3) After the single-channel mask M of size w×v is expanded to M′ of size w×v×C, it is combined with f i s and Multiply the corresponding position elements of , which will result in two sets of patch features for the blending operation, namely the source and target patch features and Right now:

[0056]

[0057] Among them, ⊙ is defined as the multiplication of the elements at corresponding positions in the matrix.

[0058] In addition to making the intermediate domain features gradually approach the distribution characteristics of the target domain, IDGM-SNet also needs to have a certain degree of discriminability for the intermediate domain features. Generally speaking, IDGM-SNet trained on the source domain undoubtedly has better discriminability on the source domain than on the target domain. Therefore, for the zero-value area in M′ that excludes local mixing, IDGM-SNet chooses to retain the source feature f i s The corresponding area of ​​f is used to enhance the distinguishability of the intermediate domain features. i s and Multiply the corresponding positions, that is

[0059]

[0060] Get the source-preserved features that retain some source features Where I is a matrix of all 1s, see Figure 3 .

[0061] (4) In order to gradually transfer knowledge from the source to the target, a domain factor λ is introduced to adjust the source and target patch features. The degree of influence on intermediate domain features. To improve the discriminability of IDGM-SNet in the target domain, our proposed IDGM-SNet gradually adapts the classifier trained on the source domain to the target domain, even if there is a large domain gap between the source and target. Therefore, the generated intermediate domain features need to gradually transition from a distribution close to the source domain to a distribution close to the target domain. Therefore, during the training phase, as the iterations progress, the ideal λ should gradually increase from 0 to 1, which means that the intermediate domain features must exhibit different distributions at different training stages. Figure 4 The Beta(α,1) distribution diagram shows that the smaller α is, the greater the probability that λ takes a smaller value, and the smaller the probability that λ takes a larger value, and vice versa. From the above analysis, it can be seen that in the mixing process, gradually increasing α helps to increase the probability of λ taking a larger value. This not only introduces greater diversity to the intermediate domain features, but also makes them gradually transition from the source distribution to the target distribution, and smoothly bridges the two extreme domains. In the experiment, Where τ and T represent the current training iteration and the total training iteration respectively.

[0062] (5) Mixing patch features based on domain factors Right now

[0063]

[0064] in, In short, the closer λ is to 0, the Closer to the source domain features. On the contrary, λ closer to 1 means Closer to the target domain characteristics.

[0065] (6) See Figure 3 Finally, the intermediate domain feature f is obtained using the following formula i m :

[0066]

[0067] (7) For each pair of source domain and target domain features f i s , The constructed intermediate domain feature f i m , generate labels according to their components. First, calculate the mixed patch features The label of M′ is denoted as a, and according to formula (3) we can get Tagged Then calculate the source preservation feature 's label. is complementary to M′, so The proportion of the area with the element 1 in the whole is 1-a, then the source retains the characteristics The label is Finally, the intermediate domain feature f i m The derived tags are:

[0068]

[0069] (8) Calculate the overall loss function value based on the derived label The overall loss function of this invention is divided into two parts: the cross entropy loss on the source domain and the cross entropy loss on the intermediate domain:

[0070]

[0071] Where β is a balance parameter, The binary cross entropy loss function is as follows:

[0072]

[0073] in, and y represent the predicted label and true label of the sample, respectively. In summary, the loss function not only considers the discriminability of the source domain features, but also evaluates the discriminability of the intermediate domain features that gradually approach the target domain distribution features, thereby enhancing the discriminability of the target domain features.

[0074] (9) Based on the overall loss function value Train the steganalysis network.

[0075] The intermediate domain features obtained from Equation (4) and their labels obtained from Equation (5) participate in the optimization process of the loss function of Equation (6), effectively building a bridge between the two extreme domains. This bridge enables IDGM-SNet to gradually transfer knowledge from the source to the target, so that the IDGM-SNet, which gradually adapts to the target domain distribution, improves the classification performance on the target domain. In addition, IDGM-SNet optimizes the cross entropy based on the labeled source domain features and the intermediate domain features associated with the generated labels instead of the unlabeled target domain features, thereby minimizing the adverse effects of target domain feature prediction errors on model training.

[0076] The specific effects of the present invention are described in detail below:

[0077] Several experiments demonstrate the advantages of this solution in alleviating the accuracy degradation caused by CSM. The experiments are divided into two aspects: spatial domain CSM and JPEG domain CSM. Its performance is compared with some recent works, namely J-Net for spatial domain CSM and Stega-SL for JPEG domain CSM.

[0078] The experiments were conducted on three publicly available datasets: BOSSbase 1.01, Alaska, and MIRFlickr25k. BOSSbase contains 10,000 512×512 grayscale images in PGM format. Alaska contains 80,000 images in various formats, with 256×256 grayscale images in PGM format selected for this experiment. The MIRFlickr 25k dataset contains 25,000 color JPEG images of various sizes.

[0079] In the spatial domain, 10,000 images were randomly selected from each dataset and converted to PGM. The 30,000 PGM images were then resized to 256×256 using the "imresize" function in MATLAB with the "Bilinear" setting. Each set of carrier images was steganographically processed using S-UNIWARD, HILL, and MIPOD at a 0.4bpp payload to obtain the secret images. To demonstrate the effectiveness of the proposed IDGM-SNet at low payloads, each set of carrier images was steganographically processed using S-UNIWARD at a 0.2bpp payload to obtain the secret images. During the pre-training phase, each set of carrier / secret images was split into training and test sets in an 8:2 ratio. During the domain adaptation phase, the training and test sets were drawn from two different datasets with the same steganographic algorithm and payload.

[0080] For the JPEG domain, 10,000 256×256 images from BOSSbase were compressed using quality factors of 75, 85, and 95, resulting in three compressed sets, the carrier images. The carrier images were steganographically modified using J-UNIWARD and UERD with a payload of 0.4 bpnc. Similarly, to verify the effectiveness of IDGM-SNet under low payload conditions in the JPEG domain, the carrier images were steganographically modified using J-UNIWARD with a payload of 0.2 bpnc. During the pre-training phase, each set was split into a training set and a test set in an 8:2 ratio. During the domain adaptation phase, the training and test sets were derived from two BOSSbase compressed sets with the same steganographic algorithm and payload but with different quality factors (e.g., 75 or 95).

[0081] In the real world, the number of images that CSM needs to detect is usually small. Therefore, in the domain adaptation stage, 500 cover / cover pairs are randomly selected from each training set and test set to represent the source domain and target domain.

[0082] IDGM-SNet uses hyperparameters c and β set to 7 and 0.8, respectively. All experiments were conducted on Pytorch 2.0.0 and an NVIDIA 4090. During pre-training, the SRNet backbone network was implemented according to the details provided in the original SRNet paper and trained for 500 epochs. During domain adaptation, the IDGM-SNet feature extraction parameters were initialized using the SRNet parameters obtained during pre-training, and the IDGM-SNet classifier parameters were randomly initialized. Training lasted for 100 epochs. The learning rate was initialized to 0.001 and reduced to 0.0005 at epoch 80. The optimizer settings were consistent with those provided in the original SRNet paper. All IDGM-SNet parameters were updated.

[0083] Spatial Domain: As shown in Tables 1, 2, and 3, when the embedding rate is 0.4bpp, the network performance of the three steganographic algorithms, S-UNIWARD, HILL, and MIPOD, is evaluated. Compared to the related work J-Net, this scheme achieves the best performance. Specifically, in the CSM of the S-UNIWARD, HILL, and MIPOD steganographic algorithms, the performance gains achieved by J-Net are -0.50-6.80%, 0.22-5.40%, and 0.17-7.82%, respectively. The performance gains achieved by IDGM-SNet are 3.20-16.50%, 1.62-16.40%, and 3.10-12.12%, respectively, all significantly higher than those achieved by J-Net. As can be seen from Table 4, when evaluating the network performance under the S-UNIWARD steganography algorithm at an embedding rate of 0.2bpp, the performance gain brought by J-Net is -3.88-2.80%, while the performance gain brought by IDGM-SNet is 4.10-10.22%, which is also significantly better than J-Net.

[0084] Table 1 Performance comparison of steganalysis networks under S-UNIWARD 0.4bpp (%)

[0085] w / o CSM SRNet CSM SRNet J-Net IDGM-SNet Bossbase→Bossbase 93.20 Flickr 25k→Bossbase 80.90 85.00 89.50 Bossbase→Bossbase 93.20 Alaska → Bossbase 75.60 82.40 84.60 Flickr 25k → Flickr 25k 80.63 Bossbase → Flickr 25k 53.50 55.90 70.00 Flickr 25k → Flickr 25k 80.63 Alaska → Flickr 25k 68.00 70.00 71.20 Alaska→Alaska 68.80 Bossbase→Alaska 55.10 54.60 63.10 Alaska→Alaska 68.80 Flickr 25k → Alaska 59.50 59.93 65.10

[0086] Table 2 Performance comparison of steganalysis networks under HILL 0.4bpp (%)

[0087]

[0088]

[0089] Table 3 Performance comparison of steganalysis networks under MIPOD 0.4bpp (%)

[0090] w / o CSM SRNet CSM SRNet J-Net IDGM-SNet Bossbase→Bossbase 83.23 Flickr 25k→Bossbase 74.48 82.30 86.60 Bossbase→Bossbase 83.23 Alaska → Bossbase 81.30 82.70 84.40 Flickr 25k → Flickr 25k 77.18 Bossbase → Flickr 25k 62.88 63.40 71.60 Flickr 25k → Flickr 25k 77.18 Alaska → Flickr 25k 65.55 71.10 71.70 Alaska→Alaska 70.13 Bossbase→Alaska 63.90 65.70 67.90 Alaska→Alaska 70.13 Flickr 25k → Alaska 63.23 63.40 67.00

[0091] Table 4 Performance comparison of steganalysis networks under S-UNIWARD 0.2bpp (%)

[0092] w / o CSM SRNet CSM SRNet J-Net IDGM-SNet Bossbase→Bossbase 86.03 Flickr 25k→Bossbase 68.38 64.50 78.60 Bossbase→Bossbase 86.03 Alaska → Bossbase 70.40 73.20 74.50 Flickr 25k → Flickr 25k 69.15 Bossbase → Flickr 25k 51.43 53.30 61.10 Flickr 25k → Flickr 25k 69.15 Alaska → Flickr 25k 56.78 59.30 61.80 Alaska→Alaska 61.45 Bossbase→Alaska 52.10 53.10 60.00 Alaska→Alaska 61.45 Flickr 25k → Alaska 54.05 54.60 58.60

[0093] JPEG Domain: Tables 5 and 6 show that when the embedding rate is 0.4 bpnc, the network performance of the J-UNIWARD and UERD steganography algorithms is evaluated. Compared to the related work Stega-SL, this scheme achieves the best performance. Specifically, in the CSM of the J-UNIWARD and UERD steganography algorithms, Stega-SL achieves performance gains of 0.80-2.40% and 0.80-8.85%, respectively, while IDGM-SNet achieves performance gains of 10.80-30.20% and 7.60-27.35%, respectively, both significantly higher than Stega-SL. Table 7 shows that when evaluating the network performance of the J-UNIWARD steganography algorithm at an embedding rate of 0.2 bpnc, Stega-SL achieves performance gains of 1.50-4.00%, while IDGM-SNet achieves performance gains of 6.00-11.20%, also significantly outperforming Stega-SL.

[0094] Table 5 Performance comparison of steganalysis networks under J-UNIWARD 0.4bpnc (%)

[0095]

[0096]

[0097] Table 6 Performance comparison of steganalysis networks under UERD 0.4bpnc (%)

[0098] w / o CSM SRNet CSM SRNet Stega-SL IDGM-SNet QF75→75 97.15 QF85→75 50.03 51.80 74.30 QF75→75 97.15 QF95→75 50.00 52.00 57.80 QF85→85 96.55 QF75→85 50.85 59.70 78.20 QF85→85 96.55 QF95→85 50.00 53.40 68.10 QF95→95 90.90 QF75→95 50.00 50.80 57.60 QF95→95 90.90 QF85→95 50.03 51.60 65.00

[0099] Table 7 Performance comparison of steganalysis networks under J-UNIWARD 0.2bpnc (%)

[0100] w / o CSM SRNet CSM SRNet Stega-SL IDGM-SNet QF75→75 76.05 QF85→75 50.20 53.10 68.70 QF75→75 76.05 QF95→75 50.00 52.30 56.00 QF85→85 74.28 QF75→85 50.80 53.80 72.00 QF85→85 74.28 QF95→85 50.00 54.00 62.30 QF95→95 62.30 QF75→95 50.20 52.40 56.40 QF95→95 62.30 QF85→95 50.30 51.80 58.40

[0101] This paper adopts the above technical solutions and proposes an IDGM based on local feature hybrid patchup (LFMP). By generating an intermediate domain, it effectively bridges the source and target domains, thereby enhancing the generalization ability of IDGM-Snet and alleviating the degradation of target domain detection accuracy caused by CSM. This paper controls the bias of the intermediate domain by sampling domain factors from the distribution Beta(α,1), so that the generated intermediate domain gradually transfers from a distribution close to the source domain to a distribution close to the target domain. This helps to gradually transfer knowledge from the source to the target, further improving the detection accuracy of IDGM-SNet for the target domain.

[0102] Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

Claims

1. A CSM-oriented steganalysis network training method based on progressive intermediate domain, characterized by: The CSM-oriented steganalysis network IDGM-Snet based on progressive intermediate domain includes the first convolution module, the intermediate domain generation module IDGM, the second convolution module and the classifier. The first convolution module, the second convolution module and the classifier constitute the backbone SRNet of IDGM-SNet. The first convolution module receives the source domain image and the target domain image and outputs the source hidden features and the target hidden features. The intermediate domain generation module (IDGM) is used to generate local mixing positions. The intermediate domain generation module (IDGM) performs local feature mixing (LFMP) on the source hidden features and target hidden features output by the first convolution module, and uses the domain factor λ to control the mixing ratio of the local feature mixing (LFMP), thereby generating intermediate domain features that gradually approach the target domain. The local feature mixing (LFMP) is used to locally mix the features of the source domain and the target domain. The second convolution module extracts features from the source domain, target domain, and intermediate domain, and the gradually adapted classifier classifies the extracted features. The training method includes the following steps: Step 1: Set the number N s The number of labeled source domain images is N t The unlabeled target domain image is input into the steganalysis network SRNet, and the source domain is output The middle layer features Predicted Label Output target domain The middle layer features Predicted Label in, represents the i-th image in the source domain image set, represents the label of the i-th image in the source domain image set; represents the jth image in the target domain image set, 0 means the image is a carrier image, 1 means the image is a secret image; Step 2: Determine the intermediate layer features f corresponding to the source domain and the target domain i s and f j t The local mixed position of , that is, generating a binary mask M of size w×v, the binary mask M includes multiple blocks of size c×c and all elements are 1; c represents the side length of the block with element 1; Step 3: Expand the single-channel binary mask M of size w×v into an expanded mask M′ of size w×v×C, and compare the expanded mask M′ with f i s and f j t The corresponding position elements of are multiplied to obtain the source patch features for the blending operation and target patch features Where w and v represent length and width respectively, and C is the number of channels; Step 4: Introduce the domain factor λ to adjust the source patch characteristics and target patch features The degree of influence on the intermediate domain features is used to obtain the domain factor mixed patch features Right now: in, Indicates the addition of corresponding positions; Step 5: For the zero-value region in M′ that excludes local mixing, select the retained source feature f i s The corresponding area of ​​the source preservation feature To enhance the distinguishability of intermediate domain features; Step 6: Source Preservation Features Combined domain factor hybrid patch features Calculate the intermediate domain feature f i m : in, Indicates source retention characteristics; Step 7: According to the intermediate domain feature f i m Generate derived labels from the components Step 8: Calculate the overall loss function value based on the derived label Step 9, based on the overall loss function value Train the steganalysis network.

2. The CSM-oriented steganalysis network training method based on progressive intermediate domain according to claim 1, characterized in that: The steps for generating the binary mask M in step 2 are as follows: Step 2-1, prepare a random binary matrix M0 using w×v binary numbers randomly drawn from the Bernoulli distribution B(1,q): in, represents an element of M0, k1 and k2 represent the row and column of the element respectively; Step 2-2, each When it is equal to 1, a c×c square block centered at (k1, k2) is generated, and each element of the square block is set to 1, and finally a binary mask M is obtained.

3. The CSM-oriented steganalysis network training method based on progressive intermediate domain according to claim 1, characterized in that: Source patch features in step 3 and target patch features The specific formula is as follows: Among them, ⊙ represents the multiplication of the elements at corresponding positions in the matrix.

4. The CSM-oriented steganalysis network training method based on progressive intermediate domain according to claim 1, characterized in that: During training, the domain factor λ in step 4 is adjusted from the distribution Beta(α,1) and satisfies Where τ and T represent the current training iteration and the total training iterations respectively.

5. The CSM-oriented steganalysis network training method based on progressive intermediate domain according to claim 1, characterized in that: In step 5, f i s and Multiply the corresponding positions to obtain the source-retained features that retain some of the source features Where I is an all-1 matrix.

6. The CSM-oriented steganalysis network training method based on progressive intermediate domain according to claim 5, characterized in that: The specific steps of step 7 are as follows: Step 7-1, calculate the mixed patch features The label of M′ is denoted as a, and the proportion of the area with 1 elements in M′ to the whole is obtained. Tagged Specifically based on formula (3), we get: Step 7-2, calculate source preservation characteristics Tags based on The complementary properties of M′ give The proportion of the area with the element 1 in the whole is 1-a, then the source retains the characteristics The label is Step 7-3, calculate the intermediate domain feature f i m The derived tags are:

7. The CSM-oriented steganalysis network training method based on progressive intermediate domain according to claim 1, characterized in that: The overall loss function value of step 8 includes the cross entropy loss function value on the source domain and the cross entropy loss function value on the intermediate domain: Where β is a balance parameter, Represents the operation output of the first convolution module; Used to represent the operation output of the second convolution module and the classifier; is the binary cross entropy loss function, which is as follows: in, and y represent the predicted label and true label of the sample, respectively.

8. A CSM-oriented steganalysis network training device based on progressive intermediate domain, characterized by: It includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps implemented by the CSM-oriented steganalysis network training method based on progressive intermediate domain as described in any one of claims 1 to 7.

Images