Encrypted space steganography method based on cyclic generative adversarial network

Abstract

The present invention discloses an encrypted space steganography method based on a cyclic generative adversarial network. The present invention uses a cyclic generative adversarial network (CycleGAN) to encrypt and decrypt digital images and secret data without the user having to do any processing on them. This is conducive to the continuous expansion of the audience of cloud services and providing continuous competitiveness in market competition. In addition, the present invention designs a noise image based on differential expansion (DE) technology, which is characterized by having a good data payload for DE embedding while retaining recovery capabilities. Experimental results show that the embedding rate of the present invention exceeds 0.47bpp and the accuracy of obtaining secret data exceeds 91%. Security analysis shows that the present invention can effectively protect user data and prevent malicious attacks. Compared with the advanced solutions in the prior art, the present invention still has obvious advantages.

Description

Technical Field

[0001] The present invention belongs to the field of cloud-based data security, and specifically relates to an encrypted space steganography method based on a cyclic generative adversarial network. Background Art

[0002] The widespread adoption of 5G communications has spurred the development of cloud-based applications. Cloud computing offers a new data processing method that can rapidly process large amounts of data. This is of great significance for data analysis and mining, helping businesses and research institutions quickly obtain valuable information. Furthermore, cloud computing has made online services more convenient. People can access a variety of online services through the internet, including e-commerce, social media, cloud storage, and video conferencing. This allows people to conveniently engage in various activities and access them anywhere, anytime. As more and more people and businesses use cloud computing services, data security issues are becoming increasingly important. In cloud computing environments, data may be exposed to various threats, such as hacker attacks, malware, and data leaks. Therefore, protecting data security is crucial. Traditional encrypted space steganography converts image data into an encrypted space, effectively ensuring secure cloud-based data transmission and storage. However, traditional encrypted space steganography requires cloud users to independently process data and store the keys used for data encryption and decryption. This places an excessive burden on users and is clearly inconsistent with the development trend of cloud services. Summary of the Invention

[0003] The purpose of this invention is to reduce the operational burden of cloud users in encrypted space steganography technology, and proposes an encrypted space steganography method based on cyclic generative adversarial networks.

[0004] The technical solution adopted in the present invention is as follows:

[0005] A steganographic method for encrypted space based on a cyclic generative adversarial network is designed to reduce the processing and storage burden on cloud users while ensuring data security during transmission and storage. The method is characterized by the following specific steps:

[0006] S1: The image owner sends the carrier image to the image encryption network, which generates an encrypted image and transmits it to the cloud;

[0007] S2: The data hider sends the secret data to the data encryption network, which generates fully encrypted data;

[0008] S3: The data hider downloads the encrypted image from the cloud, embeds the fully encrypted data into the encrypted image using differential expansion technology, and transmits the encrypted image containing the secret data back to the cloud;

[0009] S4: Users with image processing permissions can restore the encrypted image containing secret data back to the encrypted image through differential expansion technology, and restore the carrier image through the image decryption network;

[0010] S5: Users with data processing permissions can use differential expansion technology to extract fully encrypted data from the encrypted image containing secret data, and restore the original secret data through the data decryption network;

[0011] The image encryption network and the image decryption network each use two generators in a cyclic generative adversarial network to achieve mutual conversion of image data between an unencrypted natural image domain and an encrypted image domain. The natural image training samples used in training are carrier images, while the encrypted image training samples are noise images generated from the natural image training samples through a difference expansion technique.

[0012] The data encryption network and the data decryption network respectively adopt two generators in another cyclic generative adversarial network, and realize the mutual conversion of data between the secret data domain and the fully encrypted data domain through the data encryption network and the data decryption network.

[0013] Preferably, the image encryption network and image decryption network used in S1 and S4 are obtained by training two generators in a cyclic generative adversarial network; during the training process, the natural image training samples and the encrypted image training samples are mapped to each other through the two generators, thereby realizing the mutual conversion of image data in the natural image domain and the encrypted image domain; after the training is completed, the natural image can be converted into an encrypted image through the image encryption network, and the encrypted image can be restored to a natural image through the image decryption network.

[0014] Preferably, the specific method of generating a noise image from a natural image training sample by using a difference extension technology is:

[0015] First, a carrier image is obtained as a natural image training sample. Then, all pixel pairs in the natural image training sample are traversed pair by pair. For each current pixel pair, a first pixel value is randomly generated in the range of 0 to 255. Then, based on the first pixel value, a second pixel value is randomly selected within the embeddable pixel range of difference expansion. The embeddable pixel range of difference expansion is [x / 3, x / 3+170], where x is the randomly generated first pixel value. After traversing all pixel pairs in the natural image training sample, a noise image is obtained as an encrypted image training sample. All pixel pairs of the encrypted image meet the embeddable conditions of difference expansion.

[0016] Preferably, the data encryption network and data decryption network used in S2 and S5 are obtained by training two generators in a cyclic generative adversarial network; during the training process, the secret data training samples and the fully encrypted data training samples are mapped to each other through the two generators, thereby realizing the mutual conversion of data in the secret data domain and the fully encrypted data domain; after the training is completed, the secret data can be converted into fully encrypted data through the data encryption network, and the fully encrypted data will be restored to secret data through the data decryption network.

[0017] Preferably, in S3, the data hider downloads the encrypted image from the cloud and embeds the encrypted data into the encrypted image according to the differential expansion technology. The specific method is as follows:

[0018] For a pair of pixels (x, y) in the encrypted image that meet the embeddable conditions of difference expansion, use equation (1) to calculate its mean l, and use equation (2) to calculate its difference h. Then, based on the fully encrypted data s, use equation (3) to expand the difference h. Finally, use equations (4) and (5) to obtain the expanded pixel pair (x′, y′). At this time, the pixel pair (x′, y′) is embedded with the fully encrypted data s:

[0019]

[0020] h=xy, (2)

[0021] h′=2h+s, (3)

[0022]

[0023]

[0024] After embedding all the secret data, the data hider obtains the encrypted image containing the secret data and transmits it back to the cloud.

[0025] Preferably, the specific method of restoring the carrier image in S4 is as follows:

[0026] Traverse each pixel pair (x′, y′) in the encrypted image that contains the secret data, and recover the pixel pair (x″, y″) corresponding to the pixel pair (x′, y′) in the encrypted image using equations (6) to (10):

[0027]

[0028] h′=x′-y′, (7)

[0029] h″=2h+s, (8)

[0030]

[0031]

[0032] The encrypted image is then sent to the decryption image network to obtain the carrier image.

[0033] Preferably, the specific method for restoring the original secret data in S5 is as follows:

[0034] The user with image processing permission traverses each pixel pair (x′, y′) in the encrypted image containing the secret data and recovers the fully encrypted data s from it using equations (11) and (12):

[0035] h′=x′-y′, (11)

[0036] s=LSB(h′). (12)

[0037] Afterwards, the fully encrypted data is sent to the decryption data network to obtain the original secret data.

[0038] Compared with the existing technology, the present invention uses a cyclic generative adversarial network (CycleGAN) to encrypt and decrypt digital images and secret data without the user having to do any processing on them. This is conducive to the continuous expansion of the audience of cloud services and provides continuous competitiveness in market competition. In addition, the present invention designs a noise image based on difference expansion (DE) technology, which is characterized by having a good data payload for DE embedding while retaining recovery capabilities. Finally, the present invention can apply the dropout function to produce multiple sets of image and data processing networks for different social groups on a shared application platform. Experiments show that the embedding rate of the present invention exceeds 0.47bpp and the accuracy of obtaining secret data exceeds 91%. Security analysis shows that the present invention can effectively protect user data and prevent malicious attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The framework diagram of the method proposed in the present invention;

[0040] Figure 2 Generate embeddable noise image examples;

[0041] Figure 3 Schematic diagram of the CycleGAN model;

[0042] Figure 4 Multiple groups of functionally consistent networks generated by the Dropout function;

[0043] Figure 5 Experimental results of image encryption and decryption networks;

[0044] Figure 6Experimental results of data encryption and decryption networks;

[0045] Figure 7 Embedding performance of the proposed method on the dataset. DETAILED DESCRIPTION

[0046] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0047] The specific embodiments of the present invention are described in further detail below with reference to the accompanying drawings:

[0048] With the rapid development of cloud data centers, more and more users are processing and transmitting data through the cloud. Traditional encrypted space steganography converts plaintext space image data into encrypted space, effectively ensuring the security of cloud-based data transmission and storage. However, traditional encrypted space steganography requires cloud users to undertake the task of processing data and store the keys used for data encryption and decryption. This places a huge burden on users and is obviously not in line with the development trend of cloud services. In a preferred embodiment of the present invention, an encrypted space steganography method based on a cyclic generative adversarial network is provided to reduce the processing and storage burden of cloud users while ensuring the security of data during transmission and storage. Its framework diagram is shown in the figure below. Figure 1 . In the present invention, an image encryption network, an image decryption network, a data encryption network and a data decryption network are required. Among them, the image encryption network and the image decryption network respectively adopt two generators in a cyclic generative adversarial network, and realize the mutual transformation of image data between the unencrypted natural image domain and the encrypted image domain through the image encryption network and the image decryption network; and the natural image training samples used in training are carrier images, and the encrypted image training samples adopt noise images generated by the natural image training samples through a difference expansion technology. Similarly, the data encryption network and the data decryption network respectively adopt two generators in another cyclic generative adversarial network, and realize the mutual transformation of data between the secret data domain and the fully encrypted data domain through the data encryption network and the data decryption network.

[0049] The specific steps of the encrypted space steganography method based on the cyclic generative adversarial network are described in detail below, including S1 to S5:

[0050] S1: The image owner sends the carrier image to the image encryption network, which generates an encrypted image and transmits it to the cloud;

[0051] In this embodiment of the present invention, the two generators in a recurrent generative adversarial network serve as the two networks for image encryption and decryption. During actual training, the generators can map natural image training samples to encrypted image training samples, enabling the conversion of image data between the natural image domain and the encrypted image domain. After training, the image encryption network can be used to convert natural images into encrypted images, while the image decryption network can be used to restore encrypted images to natural images.

[0052] It is worth noting that while using fully encrypted images as training samples can achieve image encryption and decryption functions, the generated fully encrypted images will only provide a limited payload. In order to achieve the effective payload of the encrypted image while ensuring recovery, the present invention designs a noise image based on difference expansion technology as an encrypted image training sample. Specifically, the specific method of generating a noise image from a natural image training sample using a difference expansion technology is as follows: first, a carrier image is obtained as a natural image training sample, and then all pixel pairs in the natural image training sample are traversed pair by pair. For each current pixel pair, a first pixel value is first randomly generated in the range of 0 to 255. Then, based on the first pixel value, a second pixel value is randomly selected within the embeddable pixel range of the difference expansion. The embeddable pixel range of the difference expansion is [x / 3, x / 3+170], where x is the randomly generated first pixel value. After traversing all pixel pairs in the natural image training sample, a noise image is obtained as an encrypted image training sample. All pixel pairs of this encrypted image meet the embeddable conditions of the difference expansion. For example, when the first randomly generated pixel value is 100, the second pixel value is randomly selected from the range defined by the second set of equations y=x / 3 and y=x / 3+170, as shown in the schematic diagram. Figure 2 Therefore, all pixel pairs of the encrypted image generated using the difference expansion technique are embeddable in DE. Based on this, the present invention uses a generator to map natural image training samples to generated encrypted image training samples, thereby achieving image encryption and enabling the noisy image to achieve payload while ensuring recovery.

[0053] In the method of the present invention, CycleGAN is used to achieve the conversion between two fields, as shown in the schematic diagram. Figure 3 .

[0054] The training of CycleGAN belongs to the existing technology. In this embodiment, in order to make the generated encrypted image as close as possible to the encrypted image domain, the loss function used when training the cyclic generative adversarial network corresponding to the image encryption / decryption network is as follows:

[0055]

[0056]

[0057] In this loss function, and Represents two corresponding domains (corresponding to Figure 3 in domain and domain), and and is a generator used to generate images of the corresponding domain, and is the discriminator, which is used to make the generated image closer to the target domain, while X and Y belong to and To facilitate understanding, we can regard the carrier image to be encrypted as domain, and the encrypted image is considered domain, and vice versa.

[0058] In order to make the generated restored image as close as possible to the original image domain, the loss function can be used as follows:

[0059]

[0060] In this loss function, Reducing this loss function is beneficial to maximizing the difference between the compressed restored image and the original image.

[0061] In addition, in order to ensure the security of image data, in other embodiments of the present invention, the Dropout function can also be applied to the proposed network to generate multiple groups of networks with consistent functions, but they cannot decode and interfere with each other, as shown in the schematic diagram. Figure 4 The goal of generating multiple networks with identical functionality but unable to decrypt or interfere with each other is to create distinct image encryption / decryption networks for different social groups on a shared application platform. Each social group on the platform uses a different image encryption / decryption network, ensuring security and reliability.

[0062] S2: The data hider sends the secret data to the data encryption network, which generates encrypted data.

[0063] In this invention, the two generators in another cyclic generative adversarial network serve as the data encryption and decryption networks. During training, the generators map secret data training samples to fully encrypted data training samples, enabling data conversion between the secret data domain and the fully encrypted data domain. After training, the data encryption network converts the secret data into fully encrypted data, and the data decryption network restores the fully encrypted data to secret data.

[0064] The secret data encryption and decryption network and the image encryption and decryption network have the same network structure and loss function form, as described above and will not be repeated here.

[0065] S3: The data hider downloads the encrypted image from the cloud, embeds the fully encrypted data into the encrypted image using differential expansion technology, and transmits the encrypted image containing the secret data back to the cloud:

[0066] The data hider downloads the encrypted image from the cloud and embeds the encrypted data into the encrypted image using the differential expansion technique. Specifically, for a pair of pixels (x, y) in the encrypted image that meet the embeddable conditions of differential expansion, the mean l is calculated using equation (1), and the difference h is calculated using equation (2). The difference h is then expanded based on the full encrypted data s using equation (3). Finally, the expanded pixel pair (x′, y′) is obtained using equations (4) and (5). At this time, the pixel pair (x′, y′) is embedded with the full encrypted data s:

[0067]

[0068] h=xy, (2)

[0069] h′=2h+s, (3)

[0070]

[0071]

[0072] After embedding, the data hider can transmit the encrypted image containing the secret data back to the cloud.

[0073] S4: Users with image processing permissions can restore the encrypted image containing secret data back to the encrypted image through differential expansion technology, and restore the carrier image through the image decryption network;

[0074] After downloading the encrypted image containing secret data, a user with image processing permission can traverse each pixel pair (x′, y′) in the encrypted image containing secret data and recover the pixel pair (x″, y″) corresponding to the pixel pair (x′, y′) in the original encrypted image using equations (6) to (10):

[0075]

[0076] h′=x′-y′, (7)

[0077] h″=2h+s, (8)

[0078]

[0079]

[0080] Afterwards, the original encrypted image is fed into the decryption image network to obtain the carrier image.

[0081] S5: Users with data processing permissions can use differential expansion technology to extract fully encrypted data from the encrypted image containing secret data, and restore the original secret data through the data decryption network.

[0082] After obtaining the encrypted image containing the secret data, the user with image processing permission can traverse each pixel pair (x′, y′) in the encrypted image containing the secret data and recover the fully encrypted data s from it using equations (11) and (12):

[0083] h′=x′-y′, (11)

[0084] s=LSB(h′). (12)

[0085] Wherein: LSB represents the least significant bit algorithm.

[0086] Afterwards, the fully secret data is sent to the decryption data network to obtain the secret data.

[0087] In order to more intuitively reflect the advantages of the present invention, the corresponding experimental results in the embodiments are provided below. Figure 5 As shown in the figure, the results of image encryption and decryption in the present invention are shown. The first column is the carrier image, the second column is the encrypted image generated by the image encryption network, the third column is the encrypted image embedded with secret data, and the fourth column is the restored image generated by the image decryption network. Figure 6 The figure shows the results of secret data encryption and decryption in the present invention. The first column is the original secret data, the second column is the encrypted secret data generated by the data encryption network, and the third column is the recovered data generated by the data decryption network. Finally, this embodiment also provides the embedding performance of the proposed scheme on four common image datasets. VR represents freeing up space, LM represents recording the index that can be hidden or not hidden, EC is the embedding capacity, and ER is the embedding rate.

[0088] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A cryptographic space steganography method based on a cyclic generative adversarial network is used to reduce the processing and storage burden of cloud users while ensuring data security during transmission and storage. The method is characterized by: The specific steps are as follows: S1: The image owner sends the carrier image to the image encryption network, which generates an encrypted image and transmits it to the cloud; S2: The data hider sends the secret data to the data encryption network, which generates fully encrypted data; S3: The data hider downloads the encrypted image from the cloud, embeds the fully encrypted data into the encrypted image using differential expansion technology, and transmits the encrypted image containing the secret data back to the cloud; S4: The user with image processing permission restores the encrypted image containing secret data back to the encrypted image through the differential expansion technology, and restores the carrier image through the image decryption network; S5: A user with data processing privileges uses differential expansion technology to extract the fully encrypted data from the encrypted image containing the secret data, and recovers the original secret data through the data decryption network. The image encryption network and the image decryption network respectively use two generators in a cyclic generative adversarial network to realize the mutual conversion of image data between the unencrypted natural image domain and the encrypted image domain through the image encryption network and the image decryption network; the natural image training samples used in the training are carrier images, and the encrypted image training samples are noise images generated from the natural image training samples through a difference expansion technique; The data encryption network and the data decryption network respectively adopt two generators in another cyclic generative adversarial network, and realize the mutual conversion of data between the secret data domain and the fully encrypted data domain through the data encryption network and the data decryption network.

2. The encrypted space steganography method based on cyclic generative adversarial network according to claim 1 is characterized in that: The image encryption network and image decryption network used in S1 and S4 are obtained by training two generators in a cyclic generative adversarial network; During the training process, two generators are used to map natural image training samples and encrypted image training samples to each other, realizing the mutual conversion of image data between the natural image domain and the encrypted image domain; after training, the natural image can be converted into an encrypted image through the image encryption network, and the encrypted image can be restored to a natural image through the image decryption network.

3. The encrypted space steganography method based on cyclic generative adversarial network according to claim 2 is characterized in that: The specific method of generating a noise image from a natural image training sample by using the difference expansion technology is as follows: First, a carrier image is obtained as a natural image training sample. Then, all pixel pairs in the natural image training sample are traversed pair by pair. For each current pixel pair, a first pixel value is randomly generated in the range of 0 to 255. Then, based on the first pixel value, a second pixel value is randomly selected within the embeddable pixel range of difference expansion. The embeddable pixel range of difference expansion is [x / 3, x / 3+170], where x is the randomly generated first pixel value. After traversing all pixel pairs in the natural image training sample, a noise image is obtained as an encrypted image training sample. All pixel pairs of the encrypted image meet the embeddable conditions of difference expansion.

4. The encrypted space steganography method based on cyclic generative adversarial network according to claim 2, characterized in that: The data encryption network and data decryption network used in S2 and S5 are obtained by training two generators in a cyclic generative adversarial network; During the training process, two generators are used to map secret data training samples and fully encrypted data training samples to each other, thus realizing the mutual conversion of data between the secret data domain and the fully encrypted data domain. After training, the secret data can be converted into fully encrypted data through the data encryption network, and the fully encrypted data can be restored to secret data through the data decryption network.

5. The encrypted space steganography method based on cyclic generative adversarial network according to claim 3 is characterized in that: In S3, the data hider downloads the encrypted image from the cloud and embeds the encrypted data into the encrypted image using the differential expansion technique. The specific steps are as follows: For a pair of pixels (x, y) in the encrypted image that meet the embeddable conditions of difference expansion, use equation (1) to calculate its mean l, and use equation (2) to calculate its difference h. Then, based on the fully encrypted data s, use equation (3) to expand the difference h. Finally, use equations (4) and (5) to obtain the expanded pixel pair (x′, y′). At this time, the pixel pair (x′, y′) is embedded with the fully encrypted data s: h=xy, (2) h′=2h+s, (3) After embedding all the secret data, the data hider obtains the encrypted image containing the secret data and transmits it back to the cloud.

Images