High capacity image steganography method based on frequency domain registration fusion

Abstract

The application discloses a high-capacity image steganography method based on frequency domain registration fusion. Firstly, a reference image after discrete cosine transformation and secret information are input into a reference image encoder to be encoded, and a frequency domain reference image containing the secret information is output. Secondly, a carrier image after discrete cosine transformation and the frequency domain reference image containing the secret information are subjected to image registration fusion in the frequency domain, and then are encoded through a carrier image encoder, and are subjected to inverse discrete cosine transformation to obtain a spatial domain stego image. Then, the spatial domain stego image is input into a decoder to obtain decoded secret information. Finally, a model composed of the reference image encoder, the carrier image encoder, the decoder and a discriminator is trained, and the trained discriminator is used to judge the authenticity of the spatial domain stego image. The secret information embedded by the application has a larger capacity, the generated spatial domain stego image is more in line with a probability statistical distribution, and the stego analysis resistance is higher.

Description

Technical Field

[0001] This invention belongs to the field of information hiding, and in particular relates to a high-capacity image steganography method based on frequency domain registration and fusion. Background Technology

[0002] Image steganography, also known as image information hiding technology, is a technique that uses images as carriers of information to embed secret information into them without compromising their characteristics, thereby achieving information hiding and copyright protection. The image used for information hiding is called the carrier image, and the carrier image after embedding secret information is called the coded image.

[0003] With the development of artificial intelligence, neural network technology has been widely applied to image steganography, achieving better results than traditional methods. However, most current image steganography algorithms are based on spatial domain techniques, and their exploration of image frequency domain techniques is not extensive enough.

[0004] In summary, neural network-based image steganography has achieved significant breakthroughs in the quality of generated encrypted images, the capacity of hidden secret information, and the security of encrypted images. However, some shortcomings still exist, such as insufficient utilization of the frequency domain of the carrier image and a low extraction rate of secret information. In conclusion, there is still room for improvement in the capacity of secret information and the quality of generated images in the field of image steganography. Summary of the Invention

[0005] The purpose of this invention is to address the problem of insufficient utilization of the image frequency domain in existing generative adversarial network-based image steganography techniques, and to propose a high-capacity image steganography technique based on frequency domain registration and fusion.

[0006] This invention is designed based on generative adversarial networks and includes the following steps:

[0007] Step 1. Perform a discrete cosine transform on the reference image, input the transformed reference image and the secret information into the reference image encoder for encoding, and output the frequency domain reference image containing the secret information.

[0008] Step 2. Perform Discrete Cosine Transform on the carrier image. After performing image registration and fusion in the frequency domain with the carrier image and the frequency domain reference image containing secret information, encode the image through the carrier image encoder to output the encrypted image in the frequency domain. Then, perform Inverse Discrete Cosine Transform on the encrypted image in the frequency domain to obtain the encrypted image in the spatial domain.

[0009] Step 3. Send the cryptic image in the spatial domain into the decoder to obtain the decoded secret information.

[0010] Step 4. Train the model consisting of a reference image encoder, a carrier image encoder, a decoder, and a discriminator, and test it. Use the trained discriminator to determine the authenticity of dense images in the spatial domain.

[0011] Furthermore, in step 1, a discrete cosine transform is performed on the reference image. The transformed reference image and the secret information are then input into a reference image encoder for encoding, outputting a frequency domain reference image containing the secret information. The specific steps are as follows:

[0012] The reference image after discrete cosine transform (DCT) is preprocessed and concatenated with the secret information as a tensor before being fed into a convolutional module containing two convolutional layers. The connections between the convolutional layers employ a densely connected DenseNet approach. The DCT-transformed reference image is then added to the tensor obtained after the convolutional module via residual connections to obtain a frequency domain reference image containing the secret information.

[0013] Further, in step 2, a discrete cosine transform is performed on the carrier image. The carrier image after the discrete cosine transform and the frequency domain reference image containing the secret information are then registered and fused in the frequency domain and encoded by the carrier image encoder to output the encrypted image in the frequency domain. Then, an inverse discrete cosine transform is performed on the encrypted image in the frequency domain to obtain the encrypted image in the spatial domain. The specific steps are as follows:

[0014] 2-1 The carrier image is subjected to discrete cosine transform to obtain the DCT coefficient matrix of the carrier image in the frequency domain; the frequency domain reference image containing secret information and the DCT coefficient matrix are registered and fused in the frequency domain; the registration and fusion method is implemented by the AdaIN formula.

[0015] 2-2 The registered and fused result and the DCT coefficient matrix are concatenated by tensors and then fed into the convolution module for convolution. The convolution module contains 3 convolutional layers, and the connections between the convolutional layers adopt the DenseNet method of dense connection.

[0016] 2-3 Perform discrete cosine inverse transform on the tensor output by the convolution module to obtain a dense image in the spatial domain.

[0017] Furthermore, step 3 involves feeding the condensed image in the spatial domain into the decoder for decoding to obtain the decoded secret information. The specific steps are as follows:

[0018] 3-1 Perform discrete cosine transform on dense images in the spatial domain.

[0019] 3-2 The dense image after discrete cosine transform is input into the decoder for decoding. The decoder contains four convolutional layers, and the connections between the convolutional layers adopt a dense connection DenseNet approach. The output of the decoder is the decoded secret information.

[0020] Furthermore, step 4 trains the discriminator using the carrier image, and then uses the trained discriminator to determine the authenticity of dense images in the spatial domain. The specific steps are as follows:

[0021] The carrier image is input into the discriminator for discrimination; the discriminator consists of 4 convolutional layers, and the result obtained after the last convolutional layer is convolved is adaptively averaged pooled to obtain a score, and the discrimination result is output.

[0022] The trained discriminator is used to identify dense images in the spatial domain to determine their authenticity.

[0023] Beneficial effects of this invention:

[0024] 1. This invention embeds secret information in the frequency domain, which results in a larger capacity of embedded secret information compared to other algorithms.

[0025] 2. This invention performs registration and fusion in the frequency domain. Compared with the direct tensor stitching method, the generated spatially dense image is more consistent with the probability statistical distribution and is more natural.

[0026] 3. This invention is based on generative adversarial networks, which, compared with other algorithms, generate spatially dense images with higher resistance to steganalysis. Attached Figure Description

[0027] Figure 1 This is a flowchart of the invention;

[0028] Figure 2 This is a flowchart of the encoder of the present invention;

[0029] Figure 3 This is a flowchart of the decoder of the present invention;

[0030] Figure 4 This is a flowchart of the discriminator of the present invention;

[0031] Figure 5 These are the experimental results of the present invention on downloaded body images and generated dense images with different hidden capacities. Detailed Implementation

[0032] This invention provides a high-capacity image steganography method based on frequency domain registration and fusion. The embedding and extraction of secret information are both accomplished by a trained generative adversarial network, as detailed in the following process: Figure 1 As shown, the main steps include:

[0033] Step 1: As Figure 2 As shown, a discrete cosine transform is performed on the reference image. The reference image after discrete cosine transform and the secret information are input into the reference image encoder for encoding. The output is a frequency domain reference image containing the secret information.

[0034] 1-1 A discrete cosine transform is performed on the reference image, followed by dimensionality upscaling of the transformed image to further extract image features. This reference image is used to store secret information and will later be used for frequency domain registration and fusion with the carrier image.

[0035] Let the reference image be R, the discrete cosine transform be F, and the reference image after the discrete cosine transform be R′. Then, the discrete cosine transform of the reference image is expressed by the formula:

[0036] R′=F(R)

[0037] 1-2 The reference image after discrete cosine transform is upgraded in dimension through a convolutional layer, and the tensor after the upgrade is denoted as R1.

[0038] R1 = Conv 3→32 (R′)

[0039] Where Conv 3→32 This means that a 3-dimensional tensor is mapped to a 32-dimensional tensor through convolution, while the length and width of the tensor remain unchanged.

[0040] Next, a 0-1 matrix is ​​randomly generated based on the size of the reference image; this matrix represents the secret information. The secret information is then enlarged by a factor and concatenated with the previously obtained 32-dimensional tensor before being fed into a convolutional module for convolution. This convolutional module contains two convolutional layers. Each convolutional layer has 32 convolutional kernels, and the activation function is a LeakyReLU layer. The connections between the convolutional layers use a densely connected DenseNet approach.

[0041] Let the secret information be denoted as M∈{0,1} D×W×H Let M be a tensor of size D×W×H, where each point in the tensor is a random 0 or 1. Let the output of the convolutional module be denoted as R3. Then:

[0042] R2 = Conv 32+D→32 (Concat(R1,α*M))

[0043] R3 = Conv 64+D→3 (Concat(R1,R2,α*M)

[0044] α is a hyperparameter used to control the strength of the embedded secret information. Here, Concat represents concatenating two tensors according to their dimensions to obtain a tensor with a new dimension.

[0045] By adding the output of the convolution module to the reference image after discrete cosine transform using residual connections, a frequency domain reference image containing the secret information is obtained. This frequency domain reference image containing the secret information is denoted as R. M ,but:

[0046] R M =R′+β*R3

[0047] β is a hyperparameter that controls the intensity of the output of the convolution module.

[0048] Step 2: As Figure 2 As shown, a discrete cosine transform (DCT) is performed on the carrier image. The DCT-transformed carrier image and the frequency domain reference image containing the secret information are then registered and fused in the frequency domain and encoded by the carrier image encoder to output the cryptic image in the frequency domain. Next, an inverse DCT is performed on the cryptic image in the frequency domain to obtain the cryptic image in the spatial domain. The specific steps are as follows:

[0049] 2-1 Perform discrete cosine transform on the carrier image to obtain the DCT coefficient matrix of the carrier image in the frequency domain; perform frequency domain registration and fusion of the frequency domain reference image containing secret information and the DCT coefficient matrix.

[0050] Let the carrier image be C, the discrete cosine transform be F, and the DCT coefficient matrix of the transformed carrier image in the frequency domain be C′. Then, the discrete cosine transform of the carrier image is expressed by the formula:

[0051] C′=F(C)

[0052] The frequency domain reference image R contains the secret information. M The mean and variance of the matrix are aligned to the mean and variance of the DCT coefficient matrix C′. Here, the registration and fusion method is the AdaIN formula. The resulting frequency domain reference image containing the secret information, aligned to the DCT coefficient matrix, is denoted as R. M→C Let σ represent the variance and μ represent the mean. The formula is then expressed as:

[0053]

[0054] The AdaIN formula is generally used in the field of image fusion in the spatial domain. Its principle is to align the mean and variance of the content image features to the mean and variance of the style image. According to statistical laws, the AC coefficients in the cosine domain of an image follow a Laplace distribution, which is similar to a Gaussian distribution. Therefore, AdaIN can also be applied to registration and fusion in the frequency domain. In this method, the mean and variance of the frequency domain reference image containing the secret information are aligned to the mean and variance of the carrier image in the frequency domain. The idea of ​​image fusion is used here to improve upon previous image steganography algorithms that directly concatenate the secret information after the carrier image features, which did not consider the statistical characteristics of the carrier image and the secret information. Therefore, this method incorporates a reference image for frequency domain registration and fusion.

[0055] 2-2 The registered and fused result and the DCT coefficient matrix are concatenated using tensors and then fed into a convolutional module for convolution. This convolutional module contains three convolutional layers. Each convolutional layer has 32 convolutional kernels, and the activation function is a LeakyReLU layer. The connections between the convolutional layers use a densely connected DenseNet approach.

[0056] Let the output of the convolutional layer be denoted as E. n＝1.2.3 ,but:

[0057] E1 = Conv 6→32 (Concat(C′,R M→C ))

[0058] E2 = Conv 32→32 (E1)

[0059] E3 = Conv 64→3 (Concat(E1,E2))

[0060] In this invention, to obtain higher quality dense images, two frequency domain registration and fusion operations are performed. Specifically, the E3 obtained above is used as a new frequency domain reference image containing secret information and the DCT coefficient matrix for further registration and fusion.

[0061] 2-3 Performing an inverse discrete cosine transform on the tensor output by the convolution module yields a dense image S in the spatial domain. Let the inverse discrete cosine transform be denoted as in_F, and the encoder as E(C,R,M). The encoder can then be represented as:

[0062] E(C,R,M)=S=in_F(E3)

[0063] Step 3 as follows Figure 3 As shown, the cryptic image in the spatial domain is fed into the decoder to obtain the secret information. The specific steps are as follows:

[0064] 3-1 Performing Discrete Cosine Transform on a Compact Image in the Spatial Domain. Let the compact image in the spatial domain be S, the discrete cosine transform be F, and the compact image in the frequency domain be S′. Then:

[0065] S′=F(S)

[0066] 3-2 The dense image after discrete cosine transform is input into the decoder for decoding. The decoder contains four convolutional layers, each containing 32 convolutional kernels, and the activation function layer is a LeakyReLU layer. The connections between the convolutional layers adopt a dense connection DenseNet approach, and the output of the decoder is the decoded secret information.

[0067] Let the decoder be denoted as DE(S), and the secret information output by the decoder be denoted as M′, D n＝1.2.3Let be the output between convolutional layers. Then the formula can be written as:

[0068] D1 = Conv 3→32 (S′)

[0069] D2 = Conv 32→32 (D1)

[0070] D3 = Conv 64→32 (Concat(D1,D2))

[0071] DE(S) = M′ = Conv 96→ D(Concat(D1,D2,D3))

[0072] Step 4 as follows Figure 4 As shown, the discriminator is trained using a carrier image, and the trained discriminator is used to determine the authenticity of dense images in the spatial domain. The specific steps are as follows:

[0073] 4-1 The carrier image is input into the discriminator for discrimination; the discriminator consists of 4 convolutional layers, and the result of the last convolutional layer is adaptively averaged pooled to obtain a score. This score represents the authenticity of the image; the lower the score, the higher the authenticity of the image and the better its robustness against steganography.

[0074] 4-2 Use this discriminator to score dense images in the spatial domain.

[0075] Let the discriminator be denoted as Cr(S), Mean represent adaptive mean pooling, and the dense image in the spatial domain be denoted as S. Then the discriminator can be expressed as:

[0076] Cr(S)=Mean(Conv 32→1 (Conv 32→32 (Conv 32→32 (Conv 3→32 (S)))))

[0077] Step 5 trains the model consisting of a reference image encoder, a carrier image encoder, a decoder, and a discriminator, and then tests it. The loss function is as follows:

[0078] 5-1 uses the L2 loss function to represent the similarity between the carrier image and the spatially dense image. Let the carrier image be denoted as C, the spatially dense image as S, and the encoder loss as... but:

[0079]

[0080] 5-2 uses the cross-entropy function to represent the accuracy of the decoder in decoding the secret information. The secret information output by the decoder is denoted as M′, the original secret information as M, and the decoder loss as... but:

[0081]

[0082] 5-3 Let S be the dense image in the spatial domain, and let the discriminator loss be denoted as S0. but:

[0083]

[0084] The overall loss function of the model is shown in the following formula. The model is trained to minimize this loss function.

[0085]

[0086] Where a, b, and c are hyperparameters.

[0087] Example:

[0088] The experiments for this model were conducted using the Div2k dataset, with both the carrier and reference images randomly selected from this dataset. In the experiments, the values ​​of the two parameters α and β in the model were 100 and 40, respectively. The images in Div2k were preprocessed by randomly cropping them to 3×360×360 pixels. The model was iterated 600 times during training, at which point it tended to converge.

[0089] To conduct a control experiment, SteganoGAN was used as the control group for comparison. The main comparison parameters were the accuracy of hidden information extraction, the peak signal-to-noise ratio (PSNR) between the generated spatially dense image and the carrier image, and the structural similarity index (SSIM).

[0090] Experimental results

[0091] like Figure 5 As shown, the spatially dense image and the carrier image generated by this invention maintain a very high similarity, with both PSNR and SSIM remaining at high levels. Generally, when the PSNR is above 30, it is difficult for the human eye to perceive the difference between the two images. As the secret information capacity increases, the PSNR of this invention almost consistently remains above 40, indicating that the human eye can hardly perceive that the spatially dense image generated by this invention contains hidden information.

[0092] Table 1. Comparison of the three metrics of accuracy, PSNR, and SSIM in the experiment.

[0093]

[0094] As shown in Table 1, with the continuous increase in hidden capacity, the present invention outperforms SteganoGAN in both PSNR and SSIM. This indicates that the encoder of the present invention is superior to that of SteganoGAN. Regarding accuracy, the present invention's accuracy is slightly lower than SteganoGAN's at D=1 and D=2, but still close to 100%. Considering that in channel transmission, parity bits and error correction bits are generally designed for the transmitted hidden information, considering the specific applications in practical engineering, an accuracy close to 100% can be considered as 100% accuracy. Unlike SteganoGAN, whose accuracy drops rapidly at D=4, the present invention's accuracy only begins to decline at D=6, indicating that in practical applications, the present invention can transmit more hidden information than SteganoGAN. Theoretically, the present invention guarantees that a carrier image can hide information up to 6 times the image size, and ensures a very high similarity between the generated hidden image and the carrier image.

[0095] This invention brings prior knowledge of image steganography to deep learning-based image steganography, proposing a high-capacity image steganography method based on frequency domain registration and fusion. Experimental results show that this invention outperforms currently popular deep learning algorithms in both the realism of spatially dense images and the extraction rate of secret information, confirming that frequency domain steganography, a traditional method of image steganography, still holds a place in the current deep learning system.

Claims

1. A high-capacity image steganography method based on frequency domain registration and fusion, characterized in that, Includes the following steps: Step 1. Perform discrete cosine transform on the reference image, input the discrete cosine transformed reference image and secret information into the reference image encoder for encoding, and output the frequency domain reference image containing the secret information; The specific process of encoding with reference image encoder is as follows: the reference image after discrete cosine transform is preprocessed and concatenated with the secret information as a tensor, and then fed into the convolution module for convolution. The convolution module contains two convolutional layers, and the connection between the convolutional layers adopts the dense connection DenseNet method; the reference image after discrete cosine transform is added to the tensor obtained after passing through the convolution module through residual connection to obtain the frequency domain reference image containing the secret information. Step 2. Perform Discrete Cosine Transform (DCT) on the carrier image. Then, register and fuse the DCT-transformed carrier image and the frequency domain reference image containing the secret information in the frequency domain. Encode the image using the carrier image encoder to output the cryptic image in the frequency domain. Next, perform an inverse DCT on the cryptic image in the frequency domain to obtain the cryptic image in the spatial domain. The specific process is as follows: 2-1 Perform discrete cosine transform on the carrier image to obtain the DCT coefficient matrix of the carrier image in the frequency domain; perform frequency domain registration and fusion of the frequency domain reference image containing secret information and the DCT coefficient matrix; The registration and fusion are achieved using the AdaIN formula; 2-2 The registered and fused result and the DCT coefficient matrix are concatenated by tensors and then fed into the convolution module for convolution. The convolution module contains 3 convolutional layers, and the connection between the convolutional layers adopts the DenseNet method of dense connection. 2-3 Perform discrete cosine transform on the tensor output by the convolution module to obtain a dense image in the spatial domain; Step 3. Send the cryptic image in the spatial domain into the decoder to obtain the decoded secret information; Step 4. Train the model consisting of a reference image encoder, a carrier image encoder, a decoder, and a discriminator, and test it. Use the trained discriminator to determine the authenticity of dense images in the spatial domain.

2. The high-capacity image steganography method based on frequency domain registration and fusion according to claim 1, characterized in that, The specific process of step 3 is as follows: 3-1 Perform discrete cosine transform on dense images in the spatial domain; 3-2 The dense image after discrete cosine transform is input into the decoder for decoding; the decoder contains 4 convolutional layers, and the connection between the convolutional layers adopts the dense connection DenseNet method. The output of the decoder is the secret information obtained by decoding.

3. The high-capacity image steganography method based on frequency domain registration and fusion according to claim 2, characterized in that, The discriminator described in step 4 is composed of four convolutional layers. The result of the last convolutional layer is subjected to adaptive mean pooling to obtain a score, and the discriminant result is output.

Images