Information steganography method based on double-domain fusion generative adversarial network

Abstract

The invention relates to an information steganography method based on a double-domain fusion generative adversarial network, which belongs to the technical field of information hiding, and adopts the technical scheme that a generator part and a discriminator part are included, and a network in a generator is responsible for learning an embedded probability graph of a carrier image; an embedded simulator part of the generator converts the probability graph into an embedded cost graph; and then adding the embedded cost map and the carrier image to generate a steganographic image, finally sending the steganographic image and the carrier image into a discriminator together, and updating parameters of the generator and the discriminator through an adversarial training process. Experimental results show that compared with an existing algorithm, the steganographic image generated by the method shows better safety in the aspect of anti-steganographic analyzer detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of information hiding and provides an information steganography method based on a dual-domain fusion generative adversarial network. Background Art

[0002] With the rapid development of digitalization and cloud technology, the storage and transmission of multimedia information (text, audio, photos, videos, etc.) has become extremely convenient. However, this convenience is also accompanied by new security issues. From an individual's perspective, information security technology has become a key tool to protect personal privacy and property security from threats. Enterprises need to rely on information security technology to ensure that company secrets are not stolen and leaked. For the country, the importance of information security technology is even more prominent. Image steganography is an algorithm that embeds secret information into an image, which requires that the visual quality remains unchanged before and after embedding. In today's highly digital society, personal information and state secrets are at risk of being leaked. Therefore, the research on image steganography technology is particularly important. At present, an important direction of image steganography research is the distortion minimization framework. The goal of this research direction is to design an appropriate distortion function to guide the image embedding process to minimize the total embedding cost and obtain a higher quality steganographic image.

[0003] The HUGO algorithm was proposed by Pevny et al. in 2012. The algorithm calculates the embedding distortion of each pixel by calculating the difference in steganalysis features before and after pixel modification. The embedding distortion indicates the impact of the modification of the pixel on the security of the steganalysis. The smaller the distortion value, the greater the possibility of pixel modification. Finally, the HUGO algorithm uses Syndrome-trellis Codes (STC) under the distortion minimization framework to embed secret information. Although this method can resist the detection of Subtractive Pixel Adjacency Matrix (SPAM), it is difficult to resist the detection of Spatial Rich Model (SRM). After that, Holub et al. proposed the WOW algorithm in 2012, which defines the distortion by constructing horizontal, vertical, and degree angle directional filters and using the filter residuals of pixels in multiple directions. The larger the residual value, the more difficult it is to predict the pixel value, so the possibility of modification is greater and the distortion is smaller. The WOW algorithm significantly improves the ability to resist steganalysis. In 2013, Holub et al. proposed the UNIWARD (Universal Wavelet Relative Distortion) algorithm based on WOW, using the decomposition coefficients of directional filtering to define distortion. This method can be applied to the spatial domain and frequency domain, and the anti-detection ability is further improved. In 2015, Li et al. introduced the focus change position strategy and proposed the HILL (High-pass, Low-pass, and Low-pass) steganography method. This method uses a high-pass filter to calculate the embedding distortion of the pixel value, and then uses two low-pass filters to diffuse the distortion value, so that the pixels in similar areas are modified in the same direction, further improving the security of the steganography. With the rapid development of deep learning, image embedding cost learning can also be implemented using neural networks. In 2017, the ASDL-GAN algorithm first introduced the generative adversarial network into image steganalysis cost learning, and automatically learned the areas in the image suitable for embedding information to generate an embedding probability map. The embedding probability is then further converted into the embedding cost to guide the generation of the steganalyzer image, and is fed into the steganalyzer together with the carrier image for adversarial training. Although the algorithm takes a long time to train and is not as good as some traditional algorithms in performance, it provides ideas for the subsequent application of deep learning in image steganalyzer cost learning. Subsequently, UT-GAN was proposed, which uses the U-Net structure in the generator and improves the discriminator part. At the same time, a differentiable double-Tanh activator is introduced in the embedding simulator part. Compared with ADSL-GAN, UT-GAN has significantly improved performance and shortened training time.Huang et al. proposed Steg-GMAN, which uses a dual steganalyzer structure in the discriminator part and proposes a new training strategy to update the parameters of the generator and discriminator, so that its performance is further improved on the basis of UT-GAN. The applicant believes that in order to effectively train the generative adversarial network, the performance of the generator and the discriminator needs to be balanced. Based on this concept, we have improved the generator part proposed by Huang et al. Summary of the invention

[0004] The present invention solves the problems existing in the prior art and provides an information steganography method based on a dual-domain fusion generative adversarial network. The present invention enriches the amount of information in the model by adding a branch using wavelet convolution to extract features from another angle. At the same time, a multi-level feature fusion module is proposed at the decoding end to improve the reconstruction ability of the model.

[0005] The present invention is implemented through the following technical scheme: an information steganography method based on a dual-domain fusion generative adversarial network, including a generator part and a discriminator part, and the specific process is as follows:

[0006] Step S1, sending the carrier image to the DDFN module of the generator part, the DDFN module includes a coding layer and a decoding layer, the coding layer is a double-branch structure, and the carrier image passes through the ordinary convolution branch and the wavelet convolution branch respectively;

[0007] Step S2: Send the deep feature information learned by the encoding layer to the decoding layer. The decoding layer performs channel splicing on the two deep features obtained from the encoding layer and the upsampled features, and then sends them to the MFFD module.

[0008] Step S3: The output of the decoding layer is an embedding probability map corresponding to the carrier image. The embedding probability map and a random secret image with the same size are sent to an embedding simulator to obtain a modified map. The modified map is added to the carrier image to obtain a stego-image.

[0009] Step S4, sending the carrier image and its corresponding stego image to the discriminator part;

[0010] Step S5: Update the parameters of the generator and the discriminator according to the loss function.

[0011] Furthermore, the specific processing process of the DDFN module is as follows: the carrier image is sent to the two-branch structure of the coding layer for feature extraction, one branch uses 8 traditional convolution blocks Conv Block i, i=1,...,8; the other branch uses 7 wavelet convolutions WTConvj, j=1,...,7, and the output of Conv Block 8 is upsampled by block DeConv Block 1 and sent to the MFFD module with the outputs of Conv Block 7 and WTConv 7 for fusion; the fused output is upsampled by block DeConv Block 2 and sent to the MFFD module together with the outputs of Conv Block 6 and WTConv 6 for fusion; this operation is performed 5 times, and the output of the 7th MFFD module is upsampled by DeConv Block 8, and then processed by the Sigmoid function to output the embedding probability map.

[0012] Furthermore, the ordinary convolution block is a convolution block with two convolution kernels of size 3×3, BN normalization and Relu activation connected in sequence, and the DeConv Block is a convolution block with two convolution kernels of size 5×5, BN normalization and Relu activation connected in sequence.

[0013] Furthermore, the specific processing process of the MFFD module is as follows:

[0014] The result of concatenating the two deep features obtained from the encoding layer and the upsampled features is represented as R. Then, the features of R are processed by a 1×1 convolutional layer and LeakyReLU to obtain R1. Next, R1 is divided into two processing paths, which are described as follows:

[0015] R2=LeakyReLU(conv1(R1)) (1)

[0016] R2_1=LeakyReLU(conv3(R1)+R1) (2)

[0017] Next, we repeat the above operation twice to obtain R3, R3_1, R4 and R4_1. Feature R4_1 is processed by a 3×3 convolution layer and then by LeakyReLU to obtain R5. Finally, R2, R3, R4 and R5 are channel-concatenated and the number of channels is adjusted to match the input through a 1×1 convolution layer to obtain R6. After adding R6 to the input feature R, the final output result is obtained through LeakyReLU.

[0018] Furthermore, in step S3, a random matrix of the same size as the carrier image is used to represent the secret message, the information embedding process is simulated, and the hyperbolic tangent function is used to approximate the encoding, thereby obtaining a modified probability map. The hyperbolic tangent function is expressed as:

[0019]

[0020] where λ controls the accuracy of the model and achieves the best performance when λ=60; p i,j represents the embedding probability map, r i,j Indicates a secret message.

[0021] Furthermore, the discriminator includes two steganalyzers: Xu-Net and Yedroudj-Net, the weaker one is used to update the parameters of the discriminator, and the stronger one is used to update the parameters of the generator.

[0022] Furthermore, the loss function of the discriminator is:

[0023]

[0024] In the formula, z0 and z1 represent the correctly classified labels of the carrier image X and the steganographic image Y, respectively. i (X) and D i (Y) represents the discrimination results of X and Y input into the i-th steganalyzer respectively;

[0025] The total loss function of the generator is:

[0026]

[0027] The hyperparameters in the formula are set to: α = 10 -7 and β=1, the first part of the generator loss is:

[0028]

[0029] In the formula, H and W represent the length and width of the image, q represents the embedding load, and c is expressed as:

[0030]

[0031] In the formula represents the probability of a pixel value being modified by +1, Indicates the probability of a pixel value being modified by -1, Indicates the probability that the pixel value will not be modified;

[0032] The second part of the generator loss is:

[0033]

[0034] Formula (8) is the inverse of formula (4).

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] 1. The present invention introduces a branch of wavelet convolution to extract multiple frequency features from the image, thereby enhancing the network's ability to extract complex image features. In addition, wavelet convolution helps to expand the receptive field, enabling the network to capture contextual information more effectively.

[0037] 2. The present invention proposes a MFFD module at the decoding layer, which aims to extract various features through multi-level convolution and then fuse these features. This approach enhances the network's ability to capture details and global features, thereby enhancing its reconstruction capabilities.

[0038] 3. The present invention verifies the ability of the proposed method in improving the anti-steganalyzer detection through a series of experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is the overall flow chart of the algorithm of the present invention.

[0040] Figure 2 It is a schematic diagram of the generator part of the present invention.

[0041] Figure 3 It is a schematic diagram of the MFFD module at the decoding layer of the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0043] like Figure 1 As shown in the figure, an information steganography method based on dual-domain fusion generative adversarial network includes a generator part and a discriminator part. The generator part consists of a DDFN (Dual-Domain Fusion Network) module and an embedding simulator. The discriminator consists of two steganalyzers, XuNet and Yedroudj-Net. The weaker one is used to update the parameters of the discriminator, and the stronger one is used to update the parameters of the generator. The specific implementation steps of the proposed method are as follows:

[0044] Step S1: Send the carrier image to the DDFN module of the generator part. The specific structure of DDFN is as follows: Figure 2 As shown, the DDFN module includes a coding layer and a decoding layer. The coding layer has a dual-branch structure, and the carrier image passes through the ordinary convolution branch and the wavelet convolution branch respectively;

[0045] The ordinary convolution branch uses two convolution blocks with a convolution kernel size of 3×3, BN normalization and Relu activation, and the encoding layer uses 8 such convolution blocks; wavelet convolution is a convolution layer based on wavelet transform, which can also achieve the purpose of downsampling and also uses 7 wavelet convolution blocks.

[0046] Step S2: The deep feature information learned by the encoding layer is sent to the decoding layer. The decoding layer performs channel splicing on the two deep features obtained from the encoding layer and the upsampled features, and then sends them to the MFFD module.

[0047] The specific processing process of the DDFN module is as follows: the carrier image is sent to the two-branch structure of the coding layer for feature extraction. One branch uses 8 traditional convolution blocks Conv Block i, i=1,…,8; the other branch uses 7 wavelet convolutions WTConvj, j=1,…,7, and the output of Conv Block 8 is upsampled by DeConv Block 1 and sent to the MFFD module for fusion with the outputs of Conv Block7 and WTConv 7; the fused output is upsampled by DeConv Block2 and sent to the MFFD module together with the outputs of Conv Block 6 and WTConv 6 for fusion; this operation is repeated 5 times, and the output of the 7th MFFD module is upsampled by DeConv Block 8, and then processed by the Sigmoid function to output the embedding probability map.

[0048] The specific processing process of the MFFD module is as follows: The result of feature concatenation of the two deep features obtained from the encoding layer and the upsampled features is represented as R. Then, the features of R are processed by a 1×1 convolutional layer and LeakyReLU to obtain R1. Next, R1 is divided into two processing paths, which are described as follows:

[0049] R2=LeakyReLU(conv1(R1)) (1)

[0050] R2_1=LeakyReLU(conv3(R1)+R1) (2)

[0051] Next, we repeat the above operation twice to obtain R3, R3_1, R4 and R4_1. Feature R4_1 is processed by a 3×3 convolution layer and then by LeakyReLU to obtain R5. Finally, R2, R3, R4 and R5 are channel-concatenated and the number of channels is adjusted to match the input through a 1×1 convolution layer to obtain R6. After adding R6 to the input feature R, the final output result is obtained through LeakyReLU.

[0052] Step S3: The output of the decoding layer is an embedding probability map corresponding to the carrier image. The embedding probability map and a random secret image of the same size are sent to the embedding simulator to obtain a modified map. The modified map is added to the carrier image to obtain a stego-image.

[0053] A random matrix of the same size as the carrier image is used to represent the secret message, simulating the information embedding process, and the hyperbolic tangent function is used to approximate this encoding, thereby obtaining a modified probability map. The hyperbolic tangent function is expressed as:

[0054]

[0055] where λ controls the accuracy of the model and achieves the best performance when λ=60; p i,j represents the embedding probability map, r i,j Indicates a secret message.

[0056] Step S4: send the carrier image and its corresponding stego image to the discriminator part.

[0057] Step S5: Update the parameters of the generator and the discriminator according to the loss function.

[0058] The loss function of the discriminator is:

[0059]

[0060] In the formula, z0 and z1 represent the correctly classified labels of the carrier image X and the steganographic image Y, respectively. i (X) and D i (Y) represents the discrimination results of X and Y input into the i-th steganalyzer respectively;

[0061] The total loss function of the generator is:

[0062]

[0063] The hyperparameters in the formula are set to: α = 10 -7 and β=1, the first part of the generator loss is:

[0064]

[0065] In the formula, H and W represent the length and width of the image, q represents the embedding load, and c is expressed as:

[0066]

[0067] In the formula represents the probability of a pixel value being modified by +1, Indicates the probability of a pixel value being modified by -1, Indicates the probability that the pixel value will not be modified;

[0068] The second part of the generator loss is:

[0069]

[0070] Formula (8) is the inverse of formula (4). The model parameters of the generator and discriminator are continuously updated through the loss function to improve the performance of the network.

[0071] The system of the present invention is experimentally verified as follows

[0072] To train and evaluate the proposed model, we use the BOSSBase dataset, which consists of 10,000 images, each with a resolution of 512×512. Using MATLAB for image processing, we crop these images into 40,000 images with a resolution of 256×256 to form the BOSSBase training set. For the test set, we use the BOW2 dataset, which contains 10,000 images with a resolution of 256×256. All training and testing procedures are performed on a single NVIDIA RTX 3090 GPU, and the experimental setup is implemented using the PyTorch framework.

[0073] During the training process, we initialized the learning rate to 0.0001 and adopted a learning decay strategy, in which the learning rate decayed to 0.4 of its original value every 20 cycles. The training process was set to 72 iterations and the batch size was 24. The evaluation indicator in the present invention is the error rate. The higher the error rate, the better the resistance to the steganalyzer.

[0074] We compare the proposed method with two other advanced deep learning-based algorithms (Steg GMAN and UT-GAN) and a traditional algorithm (HILL) at different embedding capacities. In addition, we evaluate the resistance of each algorithm to steganalysis using three different steganalyzers (Xu-Net, Yedroudj-Net, and CovNet). Table 1 summarizes the experimental results, where bold text indicates the best performance in the corresponding category.

[0075] Analyzing the data in Table 1, we observe that the present invention shows the best performance under all three steganalyzers and four different embedding capacity conditions. Specifically, when facing CovNet, our method achieves a significant performance improvement of 1.33% compared to the state-of-the-art Steg-GMAN algorithm. This result shows that our method has a clear advantage in enhancing the anti-steganalysis detection of steganographic images.

[0076] Table 1 Experimental comparison of the present invention with other advanced algorithms

[0077]

[0078] Ablation experiment

[0079] To verify the effectiveness of the proposed algorithm, we designed ablation experiments. We call the unmodified baseline model "base", the version containing the wavelet convolution branch "method 1", and the version using the MFFD module "method 2". In these experiments, we use CovNet to evaluate the impact of these modifications on the performance of the algorithm. Table 2 summarizes the experimental results.

[0080] The data shown in Table 2 show that adding only the wavelet convolution branch improves the performance of the model at three different embedding capacities. However, when only the multi-level feature fusion module is introduced, the performance slightly decreases at an embedding capacity of 0.3bpp. It is worth noting that when both enhancements are applied simultaneously, the performance of the model exceeds the performance of either enhancement alone. This finding suggests that the combination of the wavelet convolution branch and the multi-level feature fusion module significantly improves the model's resistance to steganalysis.

[0081] Table 2. Quantitative evaluation of ablation experiment results

[0082]

[0083]

[0084] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An information steganography method based on dual-domain fusion generative adversarial network, characterized in that: Including the generator part and the discriminator part, the specific process is as follows: Step S1, sending the carrier image to the DDFN module of the generator part, the DDFN module includes a coding layer and a decoding layer, the coding layer is a double-branch structure, and the carrier image passes through the ordinary convolution branch and the wavelet convolution branch respectively; Step S2: Send the deep feature information learned by the encoding layer to the decoding layer. The decoding layer performs channel splicing on the two deep features obtained from the encoding layer and the upsampled features, and then sends them to the MFFD module. Step S3: The output of the decoding layer is an embedding probability map corresponding to the carrier image. The embedding probability map and a random secret image with the same size are sent to an embedding simulator to obtain a modified map. The modified map is added to the carrier image to obtain a stego-image. Step S4, sending the carrier image and its corresponding stego image to the discriminator part; Step S5: Update the parameters of the generator and the discriminator according to the loss function.

2. According to claim 1, the information steganography method based on dual-domain fusion generative adversarial network is characterized by: The specific processing process of the DDFN module is as follows: the carrier image is sent to the two-branch structure of the coding layer for feature extraction, one branch uses 8 traditional convolution blocks Conv Block i, i=1,...,8; the other branch uses 7 wavelet convolutions WTConvj, j=1,...,7, and the output of Conv Block 8 is upsampled by block DeConv Block 1 and sent to the MFFD module for fusion with the outputs of Conv Block7 and WTConv 7; the fused output is upsampled by block DeConv Block2 and sent to the MFFD module together with the outputs of Conv Block 6 and WTConv 6 for fusion; this operation is repeated 5 times, and the output of the 7th MFFD module is upsampled by DeConv Block 8, and then processed by the Sigmoid function to output the embedding probability map.

3. According to claim 2, the information steganography method based on dual-domain fusion generative adversarial network is characterized by: The ordinary convolution block is a convolution block with two convolution kernels of size 3×3, BN normalization and Relu activation connected in sequence, and the DeConv Block is a convolution block with two convolution kernels of size 5×5, BN normalization and Relu activation connected in sequence.

4. According to claim 3, the information steganography method based on dual-domain fusion generative adversarial network is characterized by: The specific processing process of the MFFD module is as follows: The result of concatenating the two deep features obtained from the encoding layer and the upsampled features is represented as R. Then, the features of R are processed by a 1×1 convolutional layer and LeakyReLU to obtain R1. Next, R1 is divided into two processing paths, which are described as follows: R2=LeakyReLU(conv1(R1)) (1) R2_1=LeakyReLU(conv3(R1)+R1) (2) Next, we repeat the above operation twice to obtain R3, R3_1, R4 and R4_1. Feature R4_1 is processed by a 3×3 convolution layer and then by LeakyReLU to obtain R5. Finally, R2, R3, R4 and R5 are channel-concatenated and the number of channels is adjusted to match the input through a 1×1 convolution layer to obtain R6. After adding R6 to the input feature R, the final output result is obtained through LeakyReLU.

5. According to claim 1, the information steganography method based on dual-domain fusion generative adversarial network is characterized by: In step S3, a random matrix of the same size as the carrier image is used to represent the secret message, simulating the information embedding process, and a hyperbolic tangent function is used to approximate the encoding, thereby obtaining a modified probability map. The hyperbolic tangent function is expressed as: where λ controls the accuracy of the model and achieves the best performance when λ=60; p i,j represents the embedding probability map, r i,j Indicates a secret message.

6. According to claim 1, the information steganography method based on dual-domain fusion generative adversarial network is characterized by: The discriminator includes two steganalyzers: Xu-Net and Yedroudj-Net, the weaker one is used to update the parameters of the discriminator, and the stronger one is used to update the parameters of the generator.

7. The information steganography method based on dual-domain fusion generative adversarial network according to claim 6 is characterized by: The loss function of the discriminator is: In the formula, z0 and z1 represent the correctly classified labels of the carrier image X and the steganographic image Y, respectively. i (X) and D i (Y) represents the discrimination results of X and Y input into the i-th steganalyzer respectively; The total loss function of the generator is: The hyperparameters in the formula are set to: α = 10 -7 and β=1, the first part of the generator loss is: In the formula, H and W represent the length and width of the image, q represents the embedding load, and c is expressed as: In the formula represents the probability of a pixel value being modified by +1, Indicates the probability of a pixel value being modified by -1, Indicates the probability that the pixel value will not be modified; The second part of the generator loss is: Formula (8) is the inverse of formula (4).

Images