Speckle-based face image steganography method

Abstract

This invention discloses a speckle-based face image steganography method, which includes two steps: image encryption and decryption. First, the transmitted light from a plaintext face image is passed through a scattering medium to generate a speckle image. This speckle image is modulated using a random phase template, and then diffracted using a Fresnel transform to generate ciphertext. Second, a GAN network is constructed, using the ciphertext as input and the plaintext as a label. The GAN network is trained using a dataset, and the ciphertext obtained in step one is input into the trained GAN network, outputting the decrypted face image. This invention's single-image encryption / decryption scheme employs a GAN network, using RD as the basic generator module and combining it with skip connections to form a U-shaped structure, avoiding the use of BN. Furthermore, feature loss is added to the loss function, improving the reconstruction quality and detail preservation of the decrypted image.

Description

Technical Field

[0001] This invention relates to the field of image encryption technology, specifically to a method for steganography of human face images based on speckle patterns. Background Technology

[0002] With the advancement of technology, facial images and videos have been widely used in social networks, smartphone unlocking, and financial transaction verification, becoming an indispensable part of the digital age. These applications not only bring convenience but also promote the digitalization and intelligentization of society. However, this is accompanied by risks to personal privacy and data security. To address this challenge, encryption technology has become one of the important means to protect personal privacy and data security. Optical cryptography systems have attracted attention in the field of facial information encryption due to their advantages of multidimensionality, large capacity, and high efficiency. However, existing encryption technologies still have the following drawbacks when applied: 1. Existing speckle-based image encryption algorithms, while maintaining simplicity and security, often suffer from poor restoration quality and loss of detail information. This is because speckle imaging is a highly random nonlinear process, and traditional decryption methods are difficult to accurately invert through explicit expressions. Furthermore, some decryption networks use Batch Normalization layers to calculate the statistics of multiple unrelated images within a Mini-Batch, resulting in weakened detail information in a single image. In addition, since the loss function only focuses on numerical performance such as pixel-level errors and ignores differences at the feature level, it further leads to the loss of high-frequency details in the decrypted image. Summary of the Invention

[0003] The purpose of this invention is to provide a speckle-based method for steganography of human face images to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a face image steganography method based on speckle, which includes two steps: image encryption and decryption;

[0005] Step 1, Image Encryption: The transmitted light from the plaintext face image is passed through a scattering medium to generate a speckle image. The speckle image is modulated using a random phase template, and the modulated image is diffracted using Fresnel transform to generate ciphertext.

[0006] Step 2, Image Decryption: Construct a GAN network with ciphertext as input and plaintext as label. Train the GAN network using the dataset. Input the ciphertext obtained in Step 1 into the trained GAN network and output the decrypted face image.

[0007] In step two, the generator of the GAN network is constructed by combining the basic module RD with a U-shaped network structure. The GAN network generator does not use batch normalization (BN) and introduces residual modules and dense modules.

[0008] Furthermore, in step one, the Fresnel transformation formula is as follows:

[0009]

[0010] in, The wavelength of light The diffraction distance. For object plane coordinates, Image plane coordinates, Let j be the light field distribution function on the object plane, and j represent the imaginary unit. It represents the differential of the x-coordinate on the object plane; It represents the differential of the y-coordinate on the object plane.

[0011] Furthermore, in step two, all activation functions in the GAN network generator are LeakyRelu(), and the loss function is:

[0012] in, , , All are hyperparameters. To generate an image and a normal image Norm loss, , The discriminator determines whether the generated image is real or fake, therefore The discriminator's judgment result is compared with 0. Norm loss, Features extracted for VGG19 pre-trained weights Norm loss.

[0013] Furthermore, the sub-modules of the basic module RD consist of three convolutional layers, with the following parameters:

[0014] Conv1:kernel_size=3,stride=1,padding=1,input_channel=inputc,out_channel=inputc

[0015] Conv2:kernel_size=3,stride=1,padding=1,input_channel=inputc*2,out_channel=inputc

[0016] Conv3:kernel_size=3,stride=1,padding=1,input_channel=inputc*3,out_channel=inputc.

[0017] Furthermore, the U-shaped network structure uses skip connections to connect certain layers of the encoder and certain layers of the decoder.

[0018] Furthermore, the generator structure of the GAN network in step two is as follows:

[0019] (1) Input a speckle image, and after one convolution layer, it becomes a 32-channel image;

[0020] (2) After another convolution, a feature map with the same size is obtained, and the number of channels becomes 64. This is then input into RD1, and RD1 outputs a feature map of size 64*128*128.

[0021] (3) Input into convolution, output a 128-channel feature map, then input into RD2, output a 128*128*128 feature map. Continue in this manner until RD4;

[0022] (4) The output of RD4 is 512 channels. After convolution, a 256*128*128 feature map is obtained. This feature map is fused with the output feature of RD3 to obtain a 512*128*128 feature map. After convolution again, a 256*128*128 feature map is obtained.

[0023] (5) The RD7 output is output after passing through two convolution layers.

[0024] Furthermore, the discriminator of the GAN network in step two is mainly divided into four parts with the following number of channels: 64, 128, 256, and 512, respectively, and finally outputs a 1*128*128 matrix.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: The single image encryption and decryption scheme of the present invention adopts a GAN network, uses RD as the basic module of the generator, combines it with skip connections to form a U-shaped structure, avoids the use of BN, and adds feature loss to the loss function, thereby improving the restoration quality and detail preservation effect of the decrypted image. Attached Figure Description

[0026] Figure 1 This is a diagram illustrating the encryption principle of images.

[0027] Figure 2 This is a basic module diagram of the generator of the GAN network in the steganography image decryption scheme of the present invention.

[0028] Figure 3 This is a generator network structure diagram of the GAN network in the steganography image decryption scheme of the present invention;

[0029] Figure 4This is a diagram of the discriminator network structure of the GAN network in the steganography image decryption scheme of the present invention;

[0030] Figure 5 This is a diagram illustrating the encryption and decryption effects of the image encryption and decryption scheme of this invention.

[0031] Figure 6 A comparison chart showing the decryption performance of GAN networks with and without Batch Normalization (BN).

[0032] Figure 7 The image shows the test results of the noise immunity of the image encryption / decryption scheme.

[0033] Figure 8 Image showing the test results of the anti-occlusion capability of the image encryption / decryption scheme;

[0034] Figure 9 for Figure 5 Histograms of pixel distribution in four plaintext images;

[0035] Figure 10 A distribution map of adjacent pixels for an image encryption / decryption scheme;

[0036] Figure 11 An image encryption scheme used in differential attacks, showing the encryption effect.

[0037] Figure 12 The decryption result of an image encryption / decryption scheme under a brute-force attack;

[0038] Figure 13 The decryption effect of an image encryption / decryption scheme under deep learning attacks;

[0039] Figure 14 This is a loss curve diagram of the network operation of the method of the present invention.

[0040] Note:

[0041] Figure 2 In the diagram, the blue line represents Denseblock and the orange line represents Residual.

[0042] Figure 5 In this diagram, (ad) represents the plaintext image, (eh) represents the corresponding ciphertext, and (il) represents the decrypted image corresponding to the ciphertext.

[0043] Figure 6 In the figure, (ad) represents the decryption effect of the GAN network of the present invention, and (eh) represents the decryption effect of the GAN network with BN.

[0044] Figure 7In the figure, (a) is the test image, (b) is the decrypted image after adding Gaussian noise with a mean of 0 and a variance of 0.5, (c) is the decrypted image after adding Gaussian noise with a mean of 0 and a variance of 1, (d) is the decrypted image after adding salt and pepper noise with a content of 0.1, and (e) is the decrypted image after adding salt and pepper noise with a content of 0.2.

[0045] Figure 8 In the image, (a) is the test image, (b) the ciphertext with the top left corner obscured by 48*48 pixels, (c) the decrypted image after obscuring the top left corner, (d) the ciphertext with the bottom right corner obscured by 48*48 pixels, and (e) the decrypted image after obscuring the bottom right corner.

[0046] Figure 9 In the diagram, (ad) is the pixel distribution histogram of the plaintext, and (eh) is the pixel distribution histogram of the ciphertext.

[0047] Figure 10 In the text, (ad) represents the distribution of adjacent pixels in plaintext, and (eh) represents the distribution of adjacent pixels in ciphertext.

[0048] Figure 11 In the image, (a) the test image Image1, (b) the image after the change, (c) the ciphertext corresponding to Image1, and (d) the ciphertext corresponding to the change.

[0049] Figure 12 middle, ;

[0050] Figure 13 In the image, (a) loss curve, (b) plaintext image 1, (c) plaintext image 2, (d) decrypted image 1, and (e) decrypted image 2;

[0051] Figure 14 In the figure, (a) is the generator loss curve, and (b) is the discriminator loss curve. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Please see the appendix Figure 1 -Appendix Figure 14 The present invention provides an embodiment of a face image steganography method based on speckle, which includes two steps: image encryption and decryption.

[0054] Step 1, Image Encryption: The transmitted light from the plaintext face image is passed through a scattering medium to generate a speckle image. The speckle image is modulated using a random phase template, and the modulated image is diffracted using Fresnel transform to generate ciphertext.

[0055] Step 2, Image Decryption: Construct a GAN network with ciphertext as input and plaintext as label. Train the GAN network using the dataset. Input the ciphertext obtained in Step 1 into the trained GAN network and output the decrypted face image.

[0056] In step two, the generator of the GAN network is constructed by combining the basic module RD with a U-shaped network structure. The GAN network generator does not use batch normalization (BN) and introduces residual modules and dense modules.

[0057] To eliminate the impact of removing Batch Normalization (BN) on the entire network, this invention introduces residual modules and dense blocks, making the solution space smoother and facilitating gradient descent (e.g., Figure 2 (As shown). Figure 2 The blue line represents the dense block, and the orange line represents the residual block. The number of input and output channels and the size of the feature map are the same for the entire basic module RD structure.

[0058] In step one of this invention, the Fresnel transformation formula is as follows:

[0059]

[0060] in, The wavelength of light The diffraction distance. For object plane coordinates, Image plane coordinates, Let j be the light field distribution function on the object plane, and j represent the imaginary unit. It represents the differential of the x-coordinate on the object plane; It represents the differential of the y-coordinate on the object plane.

[0061] In step two of this invention, all activation functions in the generator of the GAN network are LeakyRelu(), and the loss function is:

[0062] in, , , All are hyperparameters. To generate an image and a normal image Norm loss, , The discriminator determines whether the generated image is real or fake, therefore The discriminator's judgment result is compared with 0. Norm loss, Features extracted for VGG19 pre-trained weights Norm loss.

[0063] The sub-modules of the basic module RD described in this invention consist of three convolutional layers, with the following parameters:

[0064] Conv1:kernel_size=3,stride=1,padding=1,input_channel=inputc,out_channel=inputc

[0065] Conv2:kernel_size=3,stride=1,padding=1,input_channel=inputc*2,out_channel=inputc

[0066] Conv3:kernel_size=3,stride=1,padding=1,input_channel=inputc*3,out_channel=inputc.

[0067] The U-shaped network structure described in this invention uses skip connections to connect certain layers of the encoder and certain layers of the decoder (the network structure of the generator in this invention is as follows). Figure 3 (As shown). This structure helps alleviate the problems of information loss and vanishing gradients, improving the training stability and convergence speed of the network. Furthermore, due to the presence of skip connections, the U-shaped structure allows the decoder to directly access the encoder feature maps of different layers, thereby enhancing feature reusability and improving network representation capabilities and gradient propagation.

[0068] The generator structure of the GAN network in step two of this invention is as follows:

[0069] (1) Input a speckle image, and after one convolution layer, it becomes a 32-channel image;

[0070] (2) After another convolution, a feature map with the same size is obtained, and the number of channels becomes 64. This is then input into RD1, and RD1 outputs a feature map of size 64*128*128.

[0071] (3) Input into convolution, output a 128-channel feature map, then input into RD2, output a 128*128*128 feature map. Continue in this manner until RD4;

[0072] (4) The output of RD4 is 512 channels. After convolution, a 256*128*128 feature map is obtained. This feature map is fused with the output feature of RD3 to obtain a 512*128*128 feature map. After convolution again, a 256*128*128 feature map is obtained.

[0073] (5) The RD7 output is output after passing through two convolution layers.

[0074] The discriminator of the GAN network in step two of this invention is mainly divided into four parts (such as...). Figure 4 As shown in the figure, the number of channels are 64, 128, 256, and 512 respectively, and the final output is a 1*128*128 matrix.

[0075] Experimental Example 1:

[0076] To verify the image encryption and decryption effect of the steganography proposed in this invention, a simulation experiment was conducted on an RTX3090Ti. The encryption and decryption effects are as follows: Figure 5 As shown, Figure 5 (ad) represents a plaintext image. Figure 5 (eh) represents the corresponding ciphertext. As you can see, the ciphertext obscures the information in the plaintext image. Visually, the encryption is effective. Figure 5 (il) is the decrypted image corresponding to the ciphertext. It can be seen that there is no obvious difference between the plaintext image and the decrypted image. The quality of the decrypted image was quantitatively analyzed using CC and PSNR, as shown in Table 1. It can be seen that the present invention can maintain high decryption quality both visually and numerically, reflecting the feasibility of the method.

[0077] Experimental Example 2:

[0078] To verify the decryption effect of the GAN network in the steganography proposed in this invention, it was compared with a GAN network containing Batch Normalization (BN). The network was trained using the FFHQ dataset, which contains 70,000 face images, resized to 128×128, and divided into a training set of 60,000 images and a test set of 10,000 images. Figure 6 As shown, where Figure 6 (ad) This invention proposes a GAN network for decryption. Figure 6 (eh) represents the decryption performance of a GAN network with Batch Normalization (BN). Figure 6 (il) represents the decryption result from reference [1]. Figure 6 (mp) is the decryption result of reference [2]. By comparison, it can be seen that the GAN network proposed in this invention has a better decryption effect; PSNR and SSIM are shown in Table 2. The GAN network decryption method proposed in this invention has the best recovery quality and the highest structural similarity with the original image, which proves the superiority of this method.

[0079] Note: The references are as follows:

[0080] [1]Wang X, Wang W, Wei H, et al. Holographic and speckle encryption using deep learning[J]. Optics Letters, 2021, 46(23): 5794-5797.

[0081] [2]E. Denton, S. Chintala, A. Szlam, and R. Fergus. Deep generativeimage models using a laplacian pyramid of adversarial networks. In Advances in Neural Information Processing Systems (NIPS), pages 1486–1494, 2015.

[0082] Experimental Example 3:

[0083] To verify the robustness of the steganalytic image encryption and decryption scheme proposed in this invention, its noise resistance was tested, and the results are as follows: Figure 7 As shown, Figure 7 (a) is the test image. Figure 7 (b) is the decrypted image after adding Gaussian noise with a mean of 0 and a variance of 0.5. Figure 7 (c) is the decrypted image after adding Gaussian noise with a mean of 0 and a variance of 1. Figure 7 (d) is the decrypted image after adding salt-and-pepper noise with a concentration of 0.1%. Figure 7 (e) is the decrypted image after adding salt-and-pepper noise with a content of 0.2. As can be seen from the figure, the original information can still be seen in these images, and the visual effect decreases as the noise increases. The method of the present invention uses a GAN network for decryption, so after adding noise, there is no obvious noise in the decrypted image calculated by the network weights, only some blurred areas, and the pixel values ​​are relatively smooth overall. PSNR and SSIM are shown in Table 3. From the numerical values, even with a high content of noise, the decrypted image can still maintain a high quality. In summary, the image encryption and decryption scheme proposed by the present invention has a strong noise resistance capability.

[0084] Experiment Example 4:

[0085] To verify the robustness of the image encryption and decryption scheme proposed by the steganography technique of this invention, its resistance to occlusion was tested by occluding the top left and bottom right 48*48 pixels of the ciphertext respectively. The results are as follows. Figure 8 As shown in (b) and (d), the decryption effect after occlusion is as follows: Figure 8 As shown in (c) and (e), although some areas of the image become blurred, it still maintains high quality. PSNR and SSIM are shown in Table 4. After cropping 48*48 pixels at different locations, the PSNR is greater than 34dB, and the SSIM is around 0.9. Figure 8 This demonstrates that the image encryption and decryption scheme proposed by the steganography technique in this invention has strong anti-occlusion capabilities.

[0086] Experimental Example 5:

[0087] For an effective encryption algorithm, the ciphertext image should destroy the statistical features of the plaintext image. Therefore, in order to verify the security of the image encryption and decryption scheme proposed in this invention, the pixel distribution histogram of the plaintext and ciphertext is analyzed, which reflects the number of times each pixel appears in the entire image. Figure 9 (ad) is the pixel distribution histogram of the plaintext image. Figure 9 (eh) is the pixel distribution histogram of the corresponding ciphertext image. As can be seen, the ciphertext effectively disrupts the pixel distribution of the plaintext image, and the pixel values ​​of the ciphertext are slightly darker overall. In addition, the pixel distribution shapes of all ciphertext images are roughly the same, indicating that the method can map different plaintext images to a similar ciphertext space, making it impossible for attackers to crack the encryption algorithm by analyzing the correlation between the ciphertexts of different images. This demonstrates the security of the image encryption and decryption scheme proposed in this invention.

[0088] Experimental Example 6:

[0089] In the main body of an image, the pixel values ​​of adjacent pixels are very close, while those at the edges change more rapidly. Therefore, in most cases, adjacent pixels are distributed in a diagonal shape. Figure 10 As shown in (ad), this has become a common analytical technique used by attackers. Therefore, encryption algorithms need to destroy the correlation between adjacent pixels to better resist this type of attack. The image encryption and decryption scheme proposed in this invention has the following distribution of adjacent pixels in the horizontal direction: Figure 10 As shown in (eh), the points are mainly distributed in the lower left corner, and the original correlation has been disrupted, demonstrating the security of the image encryption and decryption scheme proposed in this invention.

[0090] Experiment Example 7:

[0091] Differential attacks are a common attack method that aims to break encryption systems by comparing and analyzing the correlation between plaintext and ciphertext before and after a specific change. To evaluate the anti-differential attack capability of the image encryption and decryption scheme proposed in this invention, the following test was conducted: The pixel value of the upper left corner of the test image Image1 was increased by 1 to obtain the changed image. The original image Image1, the ciphertext corresponding to Image1, the changed image, and the ciphertext corresponding to the changed image are shown below. Figure 11 As shown in Table 5, the correlation between the two ciphertext images is not visually apparent. Therefore, CC and SSIM are used to determine the correlation between the two ciphertext images. As shown in Table 5, the CC between the two ciphertext images is 0.0119 and the SSIM is 0.0692, indicating that it is difficult for attackers to analyze the correlation between them. This demonstrates that the image encryption and decryption scheme proposed in this invention can effectively resist differential attacks.

[0092] Experimental Example 8:

[0093] Exhaustive attacks, also known as brute-force attacks, systematically test all possible keys in an attempt to crack the encryption algorithm. The image encryption / decryption scheme proposed in this invention uses diffraction distance z and RPM as keys; therefore, their key sensitivity is tested separately. Higher key sensitivity indicates stronger resistance to exhaustive attacks. Figure 12 From (a) and (b), we can see that... Unable to decrypt normally when Sometimes, fuzzy decryption results can be obtained; similarly, Figure 12 (c) and (d) are respectively and The decryption result shows that each element in the RPM matrix has 10 possible values. 2 There are 10, while z has 10. 5 One possibility is that if one wants to brute-force crack the algorithm, all possible values ​​need to be enumerated. Since the computational load is quite large, the image encryption and decryption scheme proposed in this invention can effectively resist brute-force attacks.

[0094] Experiment Example 9:

[0095] Deep learning attacks are a novel attack method. Through training with a large number of plaintext-ciphertext pairs, the network is made to fit the decryption process. To evaluate the resistance of the image encryption / decryption scheme proposed in this invention to deep learning attacks, the following tests were conducted: A generative adversarial network was used to attack the algorithm proposed in this invention; the loss curve is shown below. Figure 13 As shown in (a), the loss essentially stops decreasing around 500 epochs. At this point, the input... Figure 13 After ciphering the corresponding ciphertexts (b) and (c), the output results are as follows: Figure 13As shown in (d) and (e), the differences between them are significant, and the information in the plaintext image cannot be identified. This is mainly because RPM is introduced and diffraction is performed during the encryption process, which further enhances the randomness and security of the encryption algorithm. Therefore, the image encryption and decryption scheme proposed in this invention can effectively resist deep learning attacks.

[0096] Table 1 Decryption Quality

[0097] Image1 Image2 Image3 Image4 PSNR 38.0297 37.9800 37.9371 37.9403 SSIM 0.9656 0.9412 0.9645 0.9617

[0098] Table 2 Comparison of Recovery Results

[0099]

[0100] Table 3 Decryption quality after adding noise

[0101] Between (a) and (b) Between (a) and (c) Between (a) and (d) Between (a) and (e) PSNR / dB 33.89 28.66 33.89 28.61 SSIM 0.8768 0.8248 0.8972 0.8290

[0102] Table 4. CC and PSNR of the decrypted image after occlusion

[0103] Between (a) and (c) Between (a) and (e) PSNR / dB 35.62 34.87 SSIM 0.9343 0.8987

[0104] Table 5. Correlation between the two ciphertexts

[0105] CC SSIM Between (c) and (d) 0.0119 0.0692

[0106] The network operation of this invention: The dataset used is FFHQ, containing 70,000 face images, with a normal size of 1024; 128 thumbnails are used in this paper. The dataset is randomly divided into a training set of 60,000 images and a test set of 10,000 images. Before training, the input data is normalized using the standard deviation and mean. Normalization helps the network converge and ensures that the input data follows a normal distribution, avoiding the influence of uneven data distribution. Each batch contains 2 images, and training lasts for 300 epochs. Figure 14 (a) shows the loss of the generator. Figure 14 (b) shows the discriminator's loss. The generator's loss generally decreases, indicating that the generated images are becoming increasingly similar to the labels. The discriminator's loss first decreases and then increases, suggesting that initially the discriminator could distinguish between images generated by the generator and real images. However, as the number of iterations increased and the quality of the generated images improved, the discriminator gradually became unable to make accurate judgments.

[0107] Based on the above, the advantages of this invention are as follows: It employs RD (Residual + Denseblock) as the basic module of the generator, combined with skip connections to form a U-shaped structure, allowing direct access to features from previous layers. The entire generator avoids the use of Batch Normalization (BN), thus reducing detail weakening. Feature loss is incorporated into the loss function, enabling the model to consider not only numerical performance but also the differences between features. This encryption method can be implemented optically or computer-based, exhibiting strong practicality. Experimental results show that the PSNR is higher than 37dB during normal decryption, demonstrating high robustness and security; it can effectively resist various attack methods. Therefore, this method has broad application prospects and significant practical importance in protecting image security.

[0108] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A face image steganography method based on speckle patterns, characterized in that: This method includes two steps: image encryption and decryption. Step 1, Image Encryption: The transmitted light from the plaintext face image is passed through a scattering medium to generate a speckle image. The speckle image is modulated using a random phase template, and the modulated image is diffracted using Fresnel transform to generate ciphertext. Step 2, Image Decryption: Construct a GAN network with ciphertext as input and plaintext as label. Train the GAN network using the dataset. Input the ciphertext obtained in Step 1 into the trained GAN network and output the decrypted face image. In step two, the generator of the GAN network is constructed by combining the basic module RD with a U-shaped network structure; the GAN network generator does not use batch normalization, but introduces residual modules and dense modules. In the generator of the GAN network in step two, all activation functions are LeakyRelu(), and the loss function is: ,in, , , All are hyperparameters. To generate an image and a normal image Norm loss, , The discriminator determines whether the generated image is real or fake, therefore The discriminator's judgment result is compared with 0. Norm loss, Features extracted for VGG19 pre-trained weights Norm loss; The sub-modules of the basic module RD consist of three convolutional layers, with the following parameters: Conv1:kernel_size=3,stride=1,padding=1,input_channel=inputc,out_channel=inputc Conv2:kernel_size=3,stride=1,padding=1,input_channel=inputc*2,out_channel=inputc Conv3:kernel_size=3,stride=1,padding=1,input_channel=inputc*3,out_channel=inputc; The U-shaped network structure uses skip connections to link certain layers of the encoder and certain layers of the decoder.

2. The face image steganography method based on speckle as described in claim 1, characterized in that: In step one, the Fresnel transformation formula is as follows: ， in, The wavelength of light The diffraction distance. For object plane coordinates, Image plane coordinates, Let j be the light field distribution function on the object plane, and j represent the imaginary unit. It represents the differential of the x-coordinate on the object plane; It represents the differential of the y-coordinate on the object plane.

3. The face image steganography method based on speckle as described in claim 1, characterized in that: The generator structure of the GAN network in step two is as follows: (1) Input a speckle image, and after one convolution layer, it becomes a 32-channel image; (2) After another convolution, a feature map with the same size is obtained, and the number of channels becomes 64. This is then input into RD1, and RD1 outputs a feature map of size 64*128*128. (3) Input into convolution, output a 128-channel feature map, then input into RD2, output a 128*128*128 feature map; and so on, until RD4; (4) The output of RD4 is 512 channels. After convolution, a 256*128*128 feature map is obtained. This feature map is fused with the output feature of RD3 to obtain a 512*128*128 feature map. After convolution again, a 256*128*128 feature map is obtained. (5) The RD7 output is output after passing through two convolution layers.

4. The face image steganography method based on speckle as described in claim 3, characterized in that: Step 2: The discriminator of the GAN network is mainly divided into 4 parts with 64, 128, 256 and 512 channels respectively, and finally outputs a 1*128*128 matrix.

Images