Discriminable intermediate domain training method and device for mismatch steganalysis

Abstract

The invention discloses a discriminatable intermediate domain training method and device for mismatch steganalysis, an intermediate domain network comprises a GDMR module, a DILM module, a feature extractor and a classifier, during domain adaptation, the feature extractor loads a pre-trained weight, and the classifier initializes the weight; the method comprises the following steps: extracting a feature map from a target domain image through a feature extractor, inputting the feature map into a GDMR module, carrying out maximum pooling processing on the received feature map by the GDMR module to obtain a steganography attention map, and controlling the size of a mixing region in a middle domain image through a region-level mixing ratio so as to increase the density of steganography signals in the mixing region; after the mixed region is generated, inputting the homologous domain picture and the target domain picture into a DILM module, controlling the mixing ratio of a local image by using a pixel-level mixing ratio, and finally obtaining an intermediate domain picture; and finally, jointly inputting the intermediate domain picture, the source domain picture and the target domain picture into a feature extractor and a classifier, and gradually enabling the classifier to adapt to the target domain. The intermediate domain generated by the method is more distinguishable.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a discriminable intermediate domain training method and device for mismatched steganalysis. Background Art

[0002] Image steganography refers to hiding secret data into a cover image by slightly modifying pixel values ​​or frequency coefficients; the adversary of image steganography is image steganalysis, which attempts to reveal the presence of hidden data in a suspicious image.

[0003] The development of image steganalysis has evolved from discrimination through traditional hand-crafted feature statistical differences. Traditional methods based on hand-crafted features strongly rely on a large amount of feature construction experience. With the popularity of deep learning, deep learning-based methods have eliminated the need for manual features and improved the performance of steganalysis. Therefore, a large number of deep learning-based steganalysis methods have been proposed.

[0004] Although the performance of the deep learning-based steganalysis method has achieved a very high detection rate in a laboratory environment, real life is full of uncertainties. The laboratory environment is too ideal. When the test set used by the deep learning-based steganalysis method has a different distribution from the training set, the detection rate of the test sets with different distributions usually drops significantly. In real life, the images usually detected are generally not easy to have the same distribution as the training set. Therefore, we need to overcome this problem. We call this situation carrier source mismatch (CSM).

[0005] At present, relevant methods have been proposed to solve the CSM situation. For example, ISNet uses a progressive intermediate domain as a bridge to make the classifier gradually adapt from the source domain to the target domain. ISNet relies on random sampling of the Bernoulli distribution to generate the mixed area of ​​the intermediate domain. When the randomly sampled points are expanded to square areas with the same side length, this method ignores the characteristic of steganography in embedding hidden signals in the image content. The currently popular adaptive steganography will embed hidden information into areas with complex image textures, while other areas with simple textures have fewer or no steganographic signals. The random sampling method is likely to lose areas with dense steganographic signals and easily cause the mixed area to contain a large amount of image content with no steganographic signals or few steganographic signals. This situation will result in low discriminability of the generated intermediate domain, which will have an adverse effect on the training of the classifier. Summary of the invention

[0006] The object of the present invention is to provide a discriminable intermediate domain training method and device for mismatch steganalysis.

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

[0008] A discriminable intermediate domain training method for mismatched steganalysis is proposed. The intermediate domain network includes a GDMR module, a DILM module, a feature extractor G(·) and a classifier C(·). The feature extractor G(·) and the classifier C(·) form a backbone network SRNet. During domain adaptation, the feature extractor G(·) loads pre-trained weights, and the classifier C(·) initializes weights. The training method is as follows: the target domain image is first extracted by the feature extractor G(·) to obtain a feature map, which is then input into the GDMR module. The GDMR module performs maximum pooling on the received feature map to obtain a steganalysis attention map, and controls the size of the mixed region in the intermediate domain image by the region-level mixing ratio λ, so as to increase the density of the steganalysis signal in the mixed region. After the mixed region is generated, the homologous domain image and the target domain image are input into the DILM module, and the pixel-level mixing ratio γ is used to control the mixing ratio of the local image, and finally the intermediate domain image is obtained. Finally, the intermediate domain image, the source domain image and the target domain image are input into the feature extractor G(·) and the classifier C(·) to gradually adapt the classifier to the target domain.

[0009] Furthermore, the training method comprises the following steps:

[0010] Step 1: Set the quantity to B s The labeled source domain images and B t The unlabeled target domain image is input into the steganalysis network SRNet, and the source domain image is output The feature map f s i and predicted labels And the feature map f of the target domain image t j and predicted labels in Indicates the quantity is B s The i-th picture in the source domain image of represents the label of the i-th image in the source domain image, Indicates the quantity is B t The jth image in the target domain image; 0 represents the carrier image, 1 represents the secret image;

[0011] Step 2: Transform the target domain feature map f t j Input into GDMR to obtain the steganalysis attention map pass The region-level mixing ratio λ generates a feature-level binary mask with the same size as the feature map Finally, the feature-level binary mask Upsampled to image-level binary mask O j , binary mask O jContains K regions of the same size and whose element values ​​are all 1;

[0012] Step 3: Binary mask O j , Source domain pictures and the target domain image At the same time, input it into the DILM module, and j and and Element-by-element multiplication obtains the area where the source domain and the target domain need to be mixed, and 1-O j and Multiply element by element to get the retained area of ​​the source domain image; 1 represents a matrix with all element values ​​1 that is the same size as the image;

[0013] Step 4: Use the pixel-level mixing ratio γ as a supplement to the regional mixing ratio λ, and adjust the region of the target domain image for the mixing operation according to the size of γ to generate the intermediate domain image impact.

[0014] Step 5: Add the retained area of ​​the source domain image to the mixed area blended with γ to obtain a complete intermediate domain image

[0015] Step 6: Generate the corresponding label of the middle domain according to the composition of the middle domain image

[0016] Step 7: Calculate the loss of the source domain and the intermediate domain through the cross entropy loss function, and add them together to get the total loss function;

[0017] Step 8: Train the steganalysis network through the total loss function.

[0018] Furthermore, the binary mask O of step 2 j The generation steps are as follows:

[0019] Step 2-1: Feature map f of the target domain t j Perform the maximum pooling operation on the channel to obtain the steganalysis attention map The pixel value of each position in represents the degree of attention paid by the network to the corresponding position, that is, the density of the steganographic signal in the area of ​​the image corresponding to the position;

[0020] Step 2-2: Find out The K positions with the largest pixel values ​​in the feature map, K = round(λ×h×w), h and w are the height and width of the feature map, respectively, round(·) represents the rounding operation, λ follows the Beta distribution and λ~Beta(t,1+t), where t represents the ratio of the current training number to the total training number;

[0021] Step 2-3: Create a binary mask of the same size as the feature map The element values ​​of the corresponding K positions become 1, and the other positions become 0, and the feature-level binary mask is obtained.

[0022] Step 2-4: Upsampling the feature level binary mask using the “nearest” upsampling method Upsample to get binary mask O j .

[0023] Furthermore, the steps for generating the mixed area and the reserved area in step 3 are as follows:

[0024] Step 3-1: Using binary mask O j To generate the mixed area of ​​the source domain image and the mixed area of ​​the target domain image, the formulas are and where ⊙ is element-wise multiplication;

[0025] Step 3-2: Utilize 1-O j Get the reserved area of ​​the source domain, the corresponding formula is Furthermore, γ in step 4 also obeys the Beta distribution, and γ~Beta(1-t+ε,t+ε), where ε is the Beta distribution that controls γ to maintain a monotonically decreasing trend and can effectively prevent the Beta parameter from being 0, which is set to 0.5 in the experiment;

[0026] t represents the ratio of the current training times to the total training times.

[0027] Further, in step 5, the intermediate domain image The generation steps are as follows:

[0028] Step 5-1: First, perform pixel-level blending on the mixed area. The corresponding formula is:

[0029] Step 5-2: Add the mixed area to the retained area to get the intermediate domain image The corresponding formula is

[0030] Furthermore, the label of the intermediate domain image in step 6 The generation steps are as follows:

[0031] Step 6-1: Generate the label of the mixed area first. The corresponding formula is:

[0032] Step 6-2: Calculate the label of the reserved area again. The corresponding formula is:

[0033] Step 6-3: Finally, add the labels of the two regions to get the label of the middle domain image. The formula is

[0034] Furthermore, the total loss function is the cross entropy loss of the source domain and the cross entropy loss of the intermediate domain. The total cross entropy loss is:

[0035]

[0036] Among them, B s is the number of input source domain images, ω is the weight parameter, which is set to 0.8 in the experiment, and L CE (·) is the cross entropy loss function, y and represent the true label and the predicted label respectively.

[0037] The present invention also discloses a discriminable intermediate domain training device for mismatched steganalysis, comprising a processor and a memory, wherein computer instructions are stored in the memory, 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 discriminable intermediate domain training method for mismatched steganalysis.

[0038] The present invention adopts the above technical scheme and proposes a GDMR module and a DLIM module. According to the characteristics of steganography, a steganalysis attention graph is used to find areas with dense steganalysis signals, and then λ and γ that obey the Beta distribution are used to collaboratively control the steganalysis signals in the mixed areas of the intermediate domain image to be more dense and reduce the influence of areas with fewer or no steganalysis signals on the classifier, thereby generating a more discriminable intermediate domain as a bridge to bridge the source domain and the target domain with large domain differences, so that the classifier can better adapt to the target domain, enhance the generalization ability of the network, and achieve the purpose of alleviating the CSM situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments;

[0040] Figure 1 The overall flow chart of the network for discriminable intermediate domain for mismatch steganalysis.

[0041] Figure 2It is a line graph of the average accuracy of five networks in the spatial domain under different steganography algorithms;

[0042] Figure 3 This is a line graph of the average accuracy of four networks in the JPEG domain under different steganography algorithms. Implementation

[0044] 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.

[0045] like Figures 1 to 3 As shown in FIG. 1 , the present invention discloses a discriminable intermediate domain training method for mismatched steganalysis, and proposes a high-performance steganalysis network GDNet that alleviates the CSM situation. The entire network mainly proposes how to generate a better discriminable intermediate domain to bridge the source domain and the target domain. The generation of the intermediate domain in GDNet relies on the generation of discriminable mixed regions (GDMR) module and the discriminable local image mixing (DILM) module. Among them, GDMR contains a steganalysis attention map Generation and image-level binary mask O j The generation of It is possible to find the distribution of the density of the stego signal in the image, and then use λ~Beta(t,1+t) to find the corresponding position of the area with dense stego signals, thereby generating the corresponding binary mask O j The main function of DILM is to use γ~Beta(1-t+ε,t+ε) to perform pixel-level mixing, and to supplement some of the problems left over in GDMR, thereby reducing the impact of some areas with sparse or no steganographic signals on the network, and increasing the diversity of the intermediate domain distribution. Through these two modules, a more discriminable intermediate domain is generated, and the diversity of the intermediate domain distribution is enriched, so that the classifier can better adapt to the target domain.

[0046] like Figure 1As shown in the figure, the network of discriminable intermediate domain for mismatched steganalysis includes GDMR module, DILM module, feature extractor G(·) and classifier C(·). Feature extractor G(·) and classifier C(·) constitute the backbone network SRNet, and in domain adaptation, feature extractor G(·) loads pre-trained weights, and classifier C(·) initializes weights; the training method is as follows: the target domain image is first extracted by feature extractor G(·) to obtain feature map and then input into GDMR module, GDMR module performs maximum pooling processing on the received feature map to obtain steganalysis attention map, and controls the size of mixed area in the intermediate domain image by region-level mixing ratio λ, so as to increase the density of steganalysis signal in the mixed area; after generating the mixed area, the homologous domain image and target domain image are input into DILM module, and the pixel-level mixing ratio γ is used to control the mixing ratio of local image, and finally the intermediate domain image is obtained; finally, the intermediate domain image, source domain image and target domain image are input into feature extractor G(·) and classifier C(·) together, and the classifier is gradually adapted to the target domain.

[0047] Furthermore, the training method comprises the following steps:

[0048] Step 1: Set the quantity to B s The labeled source domain images and B t The unlabeled target domain image is input into the steganalysis network SRNet, and the source domain image is output The feature map f s i and predicted labels And the feature map f of the target domain image t j and predicted labels in Indicates the quantity is B s The i-th picture in the source domain image of represents the label of the i-th image in the source domain image, Indicates the quantity is B t The jth image in the target domain image; 0 represents the carrier image, 1 represents the secret image;

[0049] Step 2: Transform the target domain feature map f t j Input into GDMR to obtain the steganalysis attention map pass The region-level mixing ratio λ generates a feature-level binary mask of the same size as the feature map Finally, the feature-level binary mask Upsampled to image-level binary mask O j , binary mask O jContains K regions of the same size and whose element values ​​are all 1;

[0050] Step 3: Binary mask O j , Source domain pictures and the target domain image At the same time, input it into the DILM module, and j and and Element-by-element multiplication obtains the area where the source domain and the target domain need to be mixed, and 1-O j and Multiply element by element to get the retained area of ​​the source domain image; 1 represents a matrix with all element values ​​1 that is the same size as the image;

[0051] Step 4: Use the pixel-level mixing ratio γ as a supplement to the regional mixing ratio λ, and adjust the region of the target domain image for the mixing operation according to the size of γ to generate the intermediate domain image impact.

[0052] Step 5: Add the retained area of ​​the source domain image to the mixed area blended with γ to obtain a complete intermediate domain image

[0053] Step 6: Generate the corresponding label of the middle domain according to the composition of the middle domain image

[0054] Step 7: Calculate the loss of the source domain and the intermediate domain through the cross entropy loss function, and add them together to get the total loss function;

[0055] Step 8: Train the steganalysis network through the total loss function.

[0056] Furthermore, the binary mask O of step 2 j The generation steps are as follows:

[0057] Step 2-1: Feature map f of the target domain t j Perform the maximum pooling operation on the channel to obtain the steganalysis attention map The pixel value of each position in represents the degree of attention paid by the network to the corresponding position, that is, the density of the steganographic signal in the area of ​​the image corresponding to the position;

[0058] Step 2-2: Find out The K positions with the largest pixel values ​​in the feature map, K = round(λ×h×w), h and w are the height and width of the feature map, respectively, round(·) represents the rounding operation, λ follows the Beta distribution and λ~Beta(t,1+t), where t represents the ratio of the current training number to the total training number;

[0059] Step 2-3: Create a binary mask of the same size as the feature map The element values ​​of the corresponding K positions become 1, and the other positions become 0, and the feature-level binary mask is obtained.

[0060] Step 2-4: Upsampling the feature level binary mask using the “nearest” upsampling method Upsample to get binary mask O j .

[0061] Furthermore, the steps for generating the mixed area and the reserved area in step 3 are as follows:

[0062] Step 3-1: Using binary mask O j To generate the mixed area of ​​the source domain image and the mixed area of ​​the target domain image, the formulas are and where ⊙ is element-wise multiplication;

[0063] Step 3-2: Utilize 1-O j Get the reserved area of ​​the source domain, the corresponding formula is

[0064] Furthermore, γ in step 4 also obeys the Beta distribution, and γ~Beta(1-t+ε,t+ε), where ε is the Beta distribution that controls γ to maintain a monotonically decreasing trend and can effectively prevent the Beta parameter from being 0, which is set to 0.5 in the experiment;

[0065] t represents the ratio of the current training times to the total training times.

[0066] Further, in step 5, the intermediate domain image The generation steps are as follows:

[0067] Step 5-1: First, perform pixel-level blending on the mixed area. The corresponding formula is:

[0068] Step 5-2: Add the mixed area to the retained area to get the intermediate domain image The corresponding formula is

[0069] Furthermore, the label of the intermediate domain image in step 6 The generation steps are as follows:

[0070] Step 6-1: Generate the label of the mixed area first. The corresponding formula is:

[0071] Step 6-2: Calculate the label of the reserved area again. The corresponding formula is:

[0072] Step 6-3: Finally, add the labels of the two regions to get the label of the middle domain image. The formula is

[0073] Furthermore, the total loss function is the cross entropy loss of the source domain and the cross entropy loss of the intermediate domain. The total cross entropy loss is:

[0074]

[0075] Among them, B s is the number of input source domain images, ω is the weight parameter, which is set to 0.8 in the experiment, and L CE (·) is the cross entropy loss function, y and represent the true label and the predicted label respectively.

[0076] The present invention also discloses a discriminable intermediate domain training device for mismatched steganalysis, comprising a processor and a memory, wherein computer instructions are stored in the memory, 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 discriminable intermediate domain training method for mismatched steganalysis.

[0077] The specific effects of the present invention are described in detail below:

[0078] In order to demonstrate the performance of the present invention in mitigating CSM, the following experiments are performed. The present invention can be performed in both the spatial domain and the JPEG domain, so it is compared with several recent popular works in these two domains, such as J-Net, RCDD, ISNet in the spatial domain, and Stega-SL and ISNet in the JPEG domain.

[0079] The experiment used three publicly available datasets, namely BOSSbase 1.01, Alaska, and MIRFlickr25k. BOSSbase contains 10,000 grayscale images in PGM format with a size of 512×512. Alaska contains 80,000 images in various formats. In the experiment, grayscale images in PGM format with a size of 256×256 were selected. The MIRFlickr 25k dataset contains 25,000 color JPEG images of different sizes.

[0080] In the spatial domain, 10,000 images were selected from the three datasets and converted into PGM format. Then, the images were resized using MATLAB and the 'imresize' function of the 'Bilinear' method was used to resize them to 256×256. At the same time, in order to verify the effectiveness under different algorithms and different payloads, these images were steganographically written using S-UNIWARD, HILL, MIPOD, and WOW at 0.4bpp, and S-UNIWARD at 0.2bpp. In the pre-training stage, the 10,000 pairs of images in each dataset were divided into training and test sets in a ratio of 8:2. In the domain adaptation stage, 500 pairs were randomly selected from the training and test sets of different datasets for CSM experiments.

[0081] In the JPEG domain, 10,000 images of Bossbase1.01 were compressed using three quality factors of 75, 85, and 95 to obtain three different data sets. As in the spatial domain, in order to verify the effectiveness of GDNet under different algorithms and different payloads, J-UNIWARD and UERD were used for steganography at 0.4bpnc, and J-UNIWARD was used for steganography at 0.2bpnc. The division of training sets and data sets in the pre-training stage and domain adaptation stage is the same as that in the spatial domain.

[0082] All experiments were implemented on Pytorch 2.0.0 and NVIDIA 4090. In the pre-training phase, the backbone network SRNet was trained, and the experimental parameters were set the same as in the original paper, and a total of 514 epochs were trained. In the domain adaptation phase, the feature extractor G(·) was initialized with pre-trained weights, but the classifier C(·) was randomly initialized with weights and only trained for 100 epochs. The settings of the optimizer were the same as in the pre-training phase, with a learning rate of 0.00025 for G(·) and 0.001 for C(·), which were maintained until the end of training.

[0083] In the spatial domain and JPEG domain, the corresponding three given data sets are used as the source domain and target domain for experiments respectively, and 6 CSM scenarios in different situations can be simulated in the two domains. Figure 2 The average accuracy of several networks under the same conditions in the spatial domain under given CSM conditions is given, where the payload of the default steganographic algorithm is 0.4 when marked as 0.2 or 0.4. It can be seen that the performance of GDNet is significantly improved under various conditions, and the effectiveness of GDNet under various conditions can be proved. Like ISNet, GDNet can not only achieve excellent performance in the spatial domain, but also has good prediction results in the JPEG domain. In the JPEG domain, we compare with ISNet and Stega-SL, where Stega-SL is designed specifically for the CSM situation in the JPEG domain. Figure 3 The experimental results of the CSM scenario of each network under the same conditions in the JPEG domain are presented. When the value is marked as 0.2 or 0.4, the payload of the default steganographic algorithm is 0.4. Similar to the performance in the spatial domain, they can surpass the previous network under various conditions and show good performance.

[0084] The present invention proposes a steganalysis attention map to represent the density distribution of the steganalysis signal, and cooperates with the effect of the steganalysis attention map through two mixing ratios sampled from the Beta distribution, so that the generated intermediate domain is more discriminable than the intermediate domain generated by the conventional method, thereby improving the generalization ability of GDNet in the target domain and alleviating the performance degradation problem caused by the CSM situation.

[0085] 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 claimed 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 of the art without making creative work are within the scope of protection of the present application.

Claims

1. A discriminative intermediate domain training method for mismatched steganalysis, characterized by: The intermediate domain network includes GDMR module, DILM module, feature extractor G(·) and classifier C(·). Feature extractor G(·) and classifier C(·) constitute the backbone network SRNet. During domain adaptation, feature extractor G(·) loads pre-trained weights, and classifier C(·) initializes weights. The training method is as follows: the target domain image is first extracted by feature extractor G(·) to obtain feature map and then input into GDMR module. GDMR module performs maximum pooling on the received feature map to obtain steganalysis attention map, and controls the size of the mixed area in the intermediate domain image by region-level mixing ratio λ to increase the density of steganalysis signal in the mixed area. After generating the mixed area, the homologous domain image and target domain image are input into DILM module, and the pixel-level mixing ratio γ is used to control the mixing ratio of local image, and finally the intermediate domain image is obtained. Finally, the intermediate domain image, source domain image and target domain image are input into feature extractor G(·) and classifier C(·) to gradually adapt the classifier to the target domain.

2. The discriminable intermediate domain training method for mismatched steganalysis according to claim 1, characterized in that: The training method includes the following steps: Step 1: Set the quantity to B s The labeled source domain images and B t The unlabeled target domain image is input into the steganalysis network SRNet, and the source domain image is output Feature map and predicted labels And the feature map of the target domain image and predicted labels in Indicates the quantity is B s The i-th picture in the source domain image of represents the label of the i-th image in the source domain image, Indicates the quantity is B t The jth image in the target domain image; 0 represents the carrier image, 1 represents the secret image; Step 2: Transform the target domain feature map Input into GDMR to obtain the steganalysis attention map pass The region-level mixing ratio λ generates a feature-level binary mask with the same size as the feature map Finally, the feature-level binary mask Upsampled to image-level binary mask O j , binary mask O j Contains K regions of the same size and whose element values ​​are all 1; Step 3: Binary mask O j , Source domain pictures and the target domain image At the same time, input it into the DILM module, and j and and Multiply element by element to get the area where the source domain and the target domain need to be mixed, and 1-O j and Multiply element by element to get the retained area of ​​the source domain image; 1 represents a matrix with all element values ​​1 that is the same size as the image; Step 4: Use the pixel-level mixing ratio γ as a supplement to the regional mixing ratio λ, and adjust the region of the target domain image for the mixing operation according to the size of γ to generate the intermediate domain image The impact of Step 5: Add the retained area of ​​the source domain image to the mixed area blended with γ to obtain a complete intermediate domain image Step 6: Generate the corresponding label of the middle domain according to the composition of the middle domain image Step 7: Calculate the loss of the source domain and the intermediate domain through the cross entropy loss function, and add them together to get the total loss function; Step 8: Train the steganalysis network through the total loss function.

3. The discriminable intermediate domain training method for mismatched steganalysis according to claim 2, characterized in that: Binary mask O of step 2 j The generation steps are as follows: Step 2-1: Feature map of the target domain Perform the maximum pooling operation on the channel to obtain the steganalysis attention map The pixel value of each position in represents the degree of attention paid by the network to the corresponding position, that is, the density of the steganographic signal in the area of ​​the image corresponding to the position; Step 2-2: Find out The K positions with the largest pixel values ​​in the feature map, K = round(λ×h×w), h and w are the height and width of the feature map, respectively, round(· represents rounding operation, λ follows Beta distribution and λ~Beta(t,1+t), where t represents the ratio of the current training number to the total training number; Step 2-3: Create a binary mask of the same size as the feature map The element values ​​of the corresponding K positions become 1, and the other positions become 0, and the feature-level binary mask is obtained. Step 2-4: Upsampling feature-level binary masks using the "nearest" upsampling method Upsample to get binary mask O j .

4. The discriminable intermediate domain training method for mismatched steganalysis according to claim 3, characterized in that: The steps for generating the mixed area and the reserved area in step 3 are as follows: Step 3-1: Using binary mask O j To generate the mixed area of ​​the source domain image and the mixed area of ​​the target domain image, the formulas are and where ⊙ is element-wise multiplication; Step 3-2: Utilize 1-O j Get the reserved area of ​​the source domain, the corresponding formula is 5. The discriminative intermediate domain training method for mismatched steganalysis according to claim 2, characterized in that: γ in step 4 also obeys the Beta distribution, and γ~Beta(1-t+ε,t+ε), where ε is the control of the Beta distribution obeyed by γ to maintain a monotonically decreasing trend and prevent the Beta parameter from being 0; t represents the ratio of the current training number to the total training number.

6. The discriminative intermediate domain training method for mismatched steganalysis according to claim 2, characterized in that: Step 5 Middle Domain Image The generation steps are as follows: Step 5-1: First, perform pixel-level blending on the mixed area. The corresponding formula is: Step 5-2: Add the mixed area to the retained area to get the intermediate domain image The corresponding formula is 7. The discriminable intermediate domain training method for mismatched steganalysis according to claim 2, characterized in that: Labels for the intermediate domain images in step 6 The generation steps are as follows: Step 6-1: Generate the label of the mixed area first. The corresponding formula is: Step 6-2: Calculate the label of the reserved area again. The corresponding formula is: Step 6-3: Finally, add the labels of the two regions to get the label of the middle domain image. The formula is 8. The discriminable intermediate domain training method for mismatched steganalysis according to claim 2, characterized in that: The total loss function is the cross entropy loss of the source domain and the cross entropy loss of the intermediate domain. The total cross entropy loss is: Among them, B s is the number of input source domain images, ω is the weight parameter, L CE (· is the cross entropy loss function, y and represent the true label and the predicted label respectively.

9. A discriminable intermediate domain training device for mismatched steganalysis, characterized in that: The device comprises a processor and a memory, wherein computer instructions are stored in the memory, 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 discriminable intermediate domain training method for mismatched steganalysis as described in any one of claims 1 to 8.

Images