Color Ghost Steganography Method Based on Multi-Discriminator Adversarial Generative Network

Abstract

The present invention discloses a color ghost steganography method based on a multi-discriminator adversarial generative network, including an encryption process: (i) a projector continuously projects a random color speckle generated by a computer onto a secret image and a non-secret image to obtain a first bucket detection value C of the secret image. sec and the second bucket detection value C of the non-secret image nsec , and use the random color speckle patterns of the two projections as the key; (ii) retain the first bucket detection value C sec and the second bucket detection value C nsec The integer parts of sec} int and {C nsec} int , where {C nsec} int As the integer part of the ciphertext, find an ε that satisfies ε{C sec} int The value is always a decimal, which is used as the decimal part of the ciphertext to obtain the final ciphertext C ste ={C nsec} int +ε{C sec} int The invention also includes the following steps: (i) reconstructing the secret image and the non-secret image of the ciphertext by using the second-order association algorithm of ghost imaging and the key; and (ii) optimizing the effect of the secret image reconstructed in the decryption process by using the adversarial generation network based on multiple discriminators to obtain a clear reconstructed image. The invention has good encryption effect and high quality of decrypted image.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to a color ghost steganography method based on a multi-discriminator adversarial generation network. Background Art

[0002] With the rapid development of modern information technology, how to solve the security problem in information transmission has become one of the important research topics. Optical information processing effectively utilizes the high efficiency and security of the optical system, and combines with information steganography technology to avoid the suspicion of information stealers. Computational ghost imaging (CGI) generates a random intensity distribution by controlling a spatial light modulator (SLM), measures the light intensity transmitted through an object using a bucket detector, and can restore the object by performing a correlation operation between the light intensity and the random intensity distribution. The correlation between the light intensity of computational ghost imaging and a randomly generated matrix by a computer provides potential applications in the field of optical information security.

[0003] To improve the security of optical encryption based on computational ghost imaging, researchers have proposed various encryption methods, including QR coding, compression encryption, multi-image encryption, visual encryption, etc. However, these applications are generally based on grayscale or binary images. Since good visual effects usually require color information, color computational ghost imaging has received great attention. In 2012, Tanha et al. proposed a color image encryption method based on computational ghost imaging by using a bucket detector with three channels (R, G, B) together with a color object. Based on a specially modulated color Hadamard matrix and Arnold transform, Qu et al. proposed a method for encrypting optical color images with a single bucket detector.

[0004] Although color computational ghost imaging has been widely applied, the sampling rate and quality problems of color computational ghost imaging have not been well solved. Summary of the Invention

[0005] The purpose of the present invention is to provide a color ghost steganography method based on a multi-discriminator adversarial generation network. The present invention can effectively perform image encryption, improve the image encryption effect, and at the same time, the decrypted image of the present invention has high visual quality.

[0006] To solve the above technical problems, the technical solution provided by the present invention is as follows: A color ghost steganography method based on a multi-discriminator adversarial generation network, including an encryption process and a decryption process. The encryption process includes the following steps:

[0007] (i) The projector continuously projects the randomly generated color speckles generated by the computer onto the secret image and the non-secret image respectively. After the reflected light is collected by the lens, the total light intensity is received by the bucket detector to obtain the first bucket detection value C of the secret image sec and the second bucket detection value C of the non-secret imagensec and use the randomly colored speckle patterns of the two projections as the key;

[0008] (ii) Retain the first bucket detection value C sec and the integer parts of the second bucket detection value C nsec are respectively denoted as {C sec} int and {C nsec} int , where {C nsec} int serves as the integer part of the ciphertext. At the same time, find an ε such that the value of ε{C sec} int is always a decimal, which serves as the fractional part of the ciphertext, thereby obtaining the final ciphertext C ste ={C nsec} int +ε{C sec} int ;

[0009] The decryption process includes the following steps:

[0010] (i) Use the ghost imaging second-order correlation algorithm and the key to reconstruct the secret image and non-secret image of the ciphertext;

[0011] (ii) Use the multi-discriminator-based adversarial generation network to optimize the effect of the secret image reconstructed during decryption to obtain a clear reconstructed image.

[0012] In the above color ghost steganography method based on the multi-discriminator adversarial generation network, the multi-discriminator-based adversarial generation network includes a generator network and two discriminator networks. The input of the generator network is the three-channel noise pattern recovered by the computational ghost imaging correlation algorithm, with a size of 64×64, followed by a deep convolutional neural network. The deep convolutional neural network consists of two 7×7 convolutional blocks, four downsampling layers, nine residual blocks, and four upsampling layers, and generates the original image through the deep convolutional neural network; PatchGAN is used in the two discriminator networks to guide the optimization of the generator network.

[0013] In the aforementioned color ghost steganography method based on the multi-discriminator adversarial generation network, a Dropout layer and a BN layer are added to the multi-discriminator-based adversarial generation network to prevent network overfitting and accelerate the convergence of the loss function.

[0014] In the aforementioned color ghost steganography method based on the multi-discriminator adversarial generation network, the fractional part ε{C sec} int adopts a random embedding method, and the embedding index is used as an additional key to improve security.

[0015] The aforementioned color ghost steganography method based on a multi-discriminator adversarial generative network. During the encryption process, for the secret image O, it is decomposed into O R , O G and O B in three parts, namely:

[0016] O = [O R O G O B ;

[0017] For the random color speckles I1, I2,..., I k ,..., I M (1 ≤ k ≤ M), each random color speckle is also decomposed into I kR , I kG and I kB in three parts. During the calculation of the total light intensity, O and I k are regarded as two sequences respectively, namely:

[0018]

[0019] where T represents the transpose operation;

[0020] Thus, the bucket detection data is obtained:

[0021] C = (C1, G2…C k , …C M )

[0022] where: C k = I kR O R + I kG O G + I kB O B .

[0023] The aforementioned color ghost steganography method based on a multi-discriminator adversarial generative network. During the decryption process, the ghost imaging second-order correlation algorithm and the key are used to reconstruct the secret image and the non-secret image. The formula is as follows:

[0024]

[0025] In the formula: is the reconstructed secret image or non-secret image.

[0026] Compared with the prior art, in the encryption process of the present invention, a computer is used to generate a series of random color speckles and project them onto a secret image and a non-secret image. The total light intensity is received by a bucket detector to obtain the corresponding secret key. Then, the first bucket detection value of the secret image is hidden into the second bucket detection value of the non-secret image to complete the encryption of the image. In the decryption process of the present invention, the received ciphertext and all secret keys can be used to reconstruct the secret image. However, due to the superposition of color Figure 3 channel information, there are serious noises and distortions in the reconstructed secret image. Therefore, in the decryption process, a multi-discriminator-based generative adversarial network is used to optimize the reconstructed color image effect, and high-visual-quality color images can be generated when the sampling rate is only 0.03. Thus, the image encryption method proposed by the present invention has the advantages of simple encryption method, good image encryption effect, high decryption efficiency, high quality of the decrypted image, and good security. In addition, the present invention can also achieve the effect of multi-image encryption in the above manner, further improving the storage capacity of information. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flowchart of the encryption and decryption processes of the present invention,

[0028] Figure 2 is a schematic diagram of color ghost steganography.

[0029] Figure 3 In (a), it is the secret image in steganography, (b) is the non-secret image, (c) is the secret image restored after decryption, and (d) is the non-secret image.

[0030] Figure 4 are the restoration effect diagrams of CGI and MDGAN when the sampling rates are set to 0.03, 0.1, and 0.2 respectively.

[0031] Figure 5 In (a) and (b), they are the secret image and non-secret image restored using incorrect random speckles during the decryption process. (c) and (d) are the secret image and non-secret image restored using incorrect secret keys during the decryption process. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The present invention will be further described below in conjunction with embodiments and drawings, but it shall not be used as a basis for limiting the present invention.

[0033] Embodiment: A color ghost steganography method based on a multi-discriminator generative adversarial network, as Figure 1 shown, includes an encryption process and a decryption process. The encryption process includes the following steps ( Figure 1 the Encryption on the left in the figure):

[0034] (i) The digital light projector (DLP) continuously projects the randomly generated color speckles produced by the computer onto the secret image and the non-secret image respectively. The reflected light is collected by the lens, and the total light intensity is received by the bucket detector to obtain the first bucket detection value C of the secret image sec and the second bucket detection value C of the non-secret image nsec , and the randomly generated color speckle patterns of the two projections are used as the key;

[0035] In this embodiment, for the secret image O, it is decomposed into O R , O G and O B in three parts, namely:

[0036] O = [O R O G O B ;

[0037] For the randomly generated color speckles I1, I2,..., I k ,..., I M (1 ≤ k ≤ M), each randomly generated color speckle is also decomposed into I kR , I kG and I kB in three parts. During the calculation of the total light intensity, O and I k are regarded as two sequences respectively, namely:

[0038]

[0039] where T represents the transpose operation;

[0040] Thus, the bucket detection data is obtained:

[0041] C = (C1, G2…C k , …C M )

[0042] where: C k = I kR O R + I kG O G + I kB O B ;

[0043] During the encryption process, the sampling rate of ghost imaging needs to be set:

[0044] δ = M / 4096

[0045] where M is the number of randomly generated speckle patterns, and 4096 is the total number of pixels of the image. The sampling rate of the non-secret image is 0.2, and the sampling rate of the secret image is 0.03.

[0046] (ii) For the steganography of the secret image and the hidden image, a specific data integration method is adopted (i.e., applying a non-obvious data integration method), such as Figure 2 shown. First, retain the integer parts of the first bucket detection value C sec and the second bucket detection value C nsec and denote them as {C sec} int and {C nsec} int respectively, where {C nsec} int is used as the integer part of the ciphertext. At the same time, find an ε such that the value of ε{C sec} int is always a decimal, which is used as the fractional part of the ciphertext. The fractional part ε{C sec} int is randomly embedded, and the embedding index is used as an additional key to improve security, thereby obtaining the final ciphertext C ste ={C nsec} int +ε{C sec} int ;

[0047] The decryption process includes the following steps (Decryption in Figure 1 ):

[0048] (i) Use the ghost imaging second-order correlation algorithm and the key to reconstruct the secret image and the non-secret image of the ciphertext; the formula is as follows:

[0049]

[0050] In the formula: is the reconstructed secret image or non-secret image.

[0051] (ii) Use the multi-discriminator-based adversarial generation network (MDGAN) to optimize the effect of the secret image reconstructed during decryption to obtain a clear reconstructed image. The multi-discriminator-based adversarial generation network includes a generator network and two discriminator networks. The input of the generator network is the three-channel noise pattern recovered by the computational ghost imaging correlation algorithm, with a size of 64×64, followed by a deep convolutional neural network. The deep convolutional neural network consists of two 7×7 convolutional blocks, four downsampling layers, nine residual blocks, and four upsampling layers, and generates the original image through the deep convolutional neural network; PatchGAN is used in the two discriminator networks to guide the optimization of the generator network. At the same time, a Dropout layer and a BN layer are added to the multi-discriminator-based adversarial generation network to prevent network overfitting and accelerate the convergence of the loss function. In this embodiment, for the comparison between the images processed during the encryption and decryption processes and the original image, see Figure 3 .Figure 3 Among them, (a) is the secret image of the present invention, (b) is the non-secret image (b), and the images recovered after correct key and network training are (c) and (d). The peak signal-to-noise ratio (PSNR) is used as the image recovery evaluation criterion. The larger the value of PSNR, the better the recovery effect. For the images (c) and (d) recovered by the present invention, their PSNR values are 18.874 and 20.0292 respectively.

[0052] For multi-image encryption, the sampling rate of the secret image is set to 0.03, and multi-image encryption can be achieved by using the above method. At this time, the additional key becomes multiple non-repeating indexes, recorded as k1, k2..., thus further improving the information storage capacity of the system.

[0053] Next, specific experiments are combined to verify the effectiveness of the present invention. First, the recovery effects of MDGAN are verified by setting the sampling rates to 20%, 10%, and 3% respectively. 100,000 unlabeled color images in the STL-10 dataset are selected as the GT images of the training set. Then, the color ghost imaging method is used to recover them as the input of the model, and 400 color ghost images corresponding to the recovered color images in the STL dataset (completely different from the training images) are selected as the test set. It is trained for 200 epochs on an RTX4000 GPU using the pytorch framework. During the training process, the initial learning rate is set to 0.0002, where the learning rate remains unchanged for the first 80 epochs and then linearly decays to 0. The batch-size is set to 64, and the Adam algorithm is used to optimize the loss function. The results of the reconstructed images are as Figure 4 shown. The peak signal-to-noise ratio (PSNR) is used as the image recovery evaluation criterion. The larger the value of PSNR, the better the recovery effect. Figure 4 The 2nd, 4th, and 6th rows of are the recovered images by the traditional correlation algorithm, and the average PSNRs obtained are 9.9574, 10.2940, and 10.3579. Figure 4 The 3rd, 5th, and 7th rows of are the recovered images by MDGAN, and the average PSNRs obtained are 17.9067, 19.5412, and 19.7710. It can be seen from this that the color images recovered by the present invention using MDGAN are significantly superior to the recovery effect of the traditional correlation algorithm.

[0054] Furthermore, the security of the present invention is investigated. A security test of incorrect key verification is carried out to verify the security of the system in the case of incorrect measurement matrices and misaligned keys. The result obtained by decrypting with incorrect random speckles here is Figure 5 (d - c), and the result obtained by decrypting with the incorrect key k is Figure 5(a-b). It can be seen from the results that the correct secret image and non-secret image cannot be restored from the incorrect random speckles. The correct non-secret image can be restored from the incorrect key, but the secret image still cannot be restored. This method verifies that the system has high security.

[0055] In summary, in the encryption process of the present invention, a series of random color speckles are generated by a computer and projected onto the secret image and the non-secret image. The total light intensity is received by a bucket detector to obtain the corresponding key; then the first bucket detection value of the secret image is hidden into the second bucket detection value of the non-secret image to complete the encryption of the image. In the decryption process of the present invention, the received ciphertext and all keys can be used to reconstruct the secret image. However, due to Figure 3 the superposition of color channel information, there are serious noises and distortions in the reconstructed secret image. Therefore, in the decryption process, an adversarial generative network based on multiple discriminators is used to optimize the reconstructed color map effect, and high-visual-quality color images can be generated when the sampling rate is only 0.03.

Claims

1. A color ghost steganography method based on a multi-discriminator adversarial generation network, characterized in that: Including an encryption process and a decryption process, the encryption process includes the following steps: (i) The projector continuously projects the random color speckles generated by the computer onto the secret image and the non-secret image respectively. After the reflected light is collected by the lens, the total light intensity is received by the bucket detector, and the first bucket detection value C of the secret image is obtained. sec And the second bucket detection value C of the non-secret image nsec , and the random color speckle patterns of the two projections are used as the key. (ii) Retain the first bucket detection value C sec and the second bucket detection value C nsec The integer parts of are respectively denoted as {C sec} int and {C nsec} int where {C nsec} int is used as the integer part of the ciphertext, and at the same time find an ε such that ε{C sec} int is always a decimal, which is used as the fractional part of the ciphertext, so as to obtain the final ciphertext C ste = {C nsec} int + ε{C sec} int ; The decryption process includes the following steps: (i) Using the ghost imaging second-order correlation algorithm and the key to reconstruct the secret image and non-secret image of the ciphertext; (ii) Using the multi-discriminator-based adversarial generation network to optimize the effect of the secret image reconstructed during decryption to obtain a clear reconstruction map.

2. The method for color ghost steganography based on a multi-discriminator adversarial generation network according to claim 1, characterized in that: The multi-discriminator-based adversarial generation network includes a generator network and two discriminator networks. The input of the generator network is a three-channel noise pattern recovered by the computational ghost imaging correlation algorithm, with a size of 64×64. Subsequently, it is a deep convolutional neural network, which consists of two 7×7 convolutional blocks, four downsampling layers, nine residual blocks, and four upsampling layers. The original image is generated through the deep convolutional neural network; PatchGAN is used in the two discriminator networks to guide the optimization of the generator network.

3. The color ghost steganography method based on the multi-discriminator adversarial generation network according to claim 2, characterized in that: A Dropout layer and a BN layer are added to the multi-discriminator-based adversarial generation network to prevent network overfitting and accelerate the convergence of the loss function.

4. The color ghost steganography method based on the multi-discriminator adversarial generation network according to claim 1, characterized in that: The fractional part ε{C of the ciphertext sec} int Adopt a random embedding method, and embed the index as an additional key to improve security.

5. The color ghost steganography method based on the multi-discriminator adversarial generation network according to claim 1, characterized in that: During the encryption process, for the secret image O, it is decomposed into O R , O G and O B in three parts, namely: O = [O R O G O B ; For the random color speckles I1, I2,..., I k ,..., I M (1 ≤ k ≤ M), each random color speckle is similarly decomposed into I kR , I kG and I kB in three parts. During the calculation of the total light intensity, O and I k are regarded as two sequences respectively, that is: Wherein, T represents the transpose operation; Thus, the bucket detection data is obtained: C = (C1, G2…C k , …C M ) Where: C k = I kR O R + I kG O G + I kB O B .

6. The color ghost steganography method based on the multi-discriminator adversarial generation network according to claim 5, characterized in that: During the decryption process, the ghost imaging second-order correlation algorithm and the key are used to reconstruct the secret image and non-secret image, and the formula is as follows: Wherein: is the reconstructed secret image or non-secret image.

Images