Steganography method based on generative image

Abstract

The invention relates to the technical field of network security, and discloses a steganography method based on a generative image, which comprises the following steps: determining secret information, a shared key and a first feature group set, and encoding the secret information to obtain a binary string; acquiring a second feature group set according to the shared key and the first feature group set; obtaining a first feature text set according to the second feature group set and the binary string; randomly selecting one feature text from the first feature text set as an input feature text, and obtaining a secret-containing image by using the input feature text and a carrier image; after iteration is finished, a final secret-containing image is obtained; and performing secret communication by using the secret-containing image. According to the method, the capacity, the safety and the robustness of generative image steganography can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of network security, and in particular to a steganographic method based on generative images. Background Art

[0002] With the digitization of information transmission and the widespread use of network media, people have put forward higher requirements for information security and privacy protection. The encryption and decryption methods of basic cryptography can no longer meet most information security needs. Therefore, steganography came into being. Different from cryptography, steganography is to directly embed secret information messages in carrier data that is not perceptually abnormal and send them to the other party. It is not easy to arouse the suspicion of attackers, and to a certain extent ensures the security of communication.

[0003] Traditional information hiding methods mainly focus on embedded steganography methods, among which images are the most commonly used carriers in embedded steganography methods. Image embedded steganography methods send secret images and transmit secret information by modifying the spatial and frequency domains of images or according to a pre-designed mapping relationship library. However, as long as the carrier is modified, traces of modification will be left, so that the steganalysis algorithm may successfully detect the existence of hidden behavior. In order to make the information hiding method fundamentally resist steganalysis, scholars in the field of information hiding have proposed the concept of Coverless Information Hiding (CIH). Compared with traditional methods, the CIH method can directly generate or retrieve stego carriers based on secret information without modification.

[0004] At present, image carrier-free information hiding algorithms are mainly divided into two categories: mapping carrier-free information hiding and generative carrier-free information hiding. The mapping carrier-free information hiding algorithm of images refers to establishing a mapping relationship between images and secret information based on the color, pixel and other features of the image itself, and then sending the image containing the secret information to achieve the information hiding process. Although the mapping image selection method directly uses natural images to transmit secret information, which can fundamentally resist the detection of various steganographic algorithms, its steganographic capacity is very limited, and it is necessary to establish a large-scale natural image database in advance to retrieve natural images that can express secret information as stego images. The generative carrier-free information hiding algorithm of images refers to realizing information hiding in the process of carrier image generation. In addition, the rise of deep learning and neural network technology has created conditions for the development of generative steganography. The emergence of models such as generative adversarial networks (GANs) has made the generation and modification of images more efficient and realistic, which provides new possibilities for embedding information into images.

[0005] Among the existing image generation steganography methods, the steganography method for face image generation is very practical. In daily life, many people often use face images as avatars for social applications. Therefore, this type of method will hardly arouse the suspicion of attackers. Some current steganography methods for face image generation are as follows: Cao Yi et al. proposed a generative image steganography method based on the generation of animated characters. First, it converts the secret information into a set of attribute labels of the animated characters, such as hairstyle, hair color, eye color, etc. Then, with the attribute label set as the constraint condition, the GANs-based animated character generation network is used to generate an image set that meets the constraint condition. After that, the quality of the generated images is evaluated, and the animated characters with better quality are selected as the steganographic images. Finally, the receiver extracts the attribute labels representing the secret information according to the corresponding method to obtain the secret information. Each secret image generated by this method can hide about 14 bits of secret information, and the steganographic capacity is low; the image itself has a low resolution, which is relatively blurry to the human eye, and it is easy to see that it is generated by AI. If it is frequently transmitted on a public channel, it is easy to arouse the suspicion of attackers; the secret image generated by this method has low robustness, and after some attacks such as noise, filtering and image compression, the accuracy of its secret information extraction will be greatly reduced.

[0006] Next, Bi Xinliang et al. proposed a generative steganography method based on the StarGAN model to generate real face images. In this scheme, a specific face dataset with style labels is first used to train the StarGAN model to achieve the effect of completing multi-style image generation. During communication, the sender encodes the secret information in binary and uses it as the target label for style conversion. Then, the target style label is input into the model to generate an image of the corresponding style and send it to the receiver. After receiving the information, the receiver uses the discriminator network to obtain the style label code, and decodes it by reverse mapping to obtain the corresponding secret information. The hidden capacity of the secret image generated by this method is subject to the style label of the training data set. That is, when the style label contains 40 different categories, the theoretical steganographic capacity of the secret image of this method is 40 bits. However, in practical applications, it is also necessary to consider whether there will be semantic contradictions between the labels, such as the conflict between the two labels "female" and "beard". This method also needs to consider that too many style labels will lead to a significant reduction in the secret information extraction rate. Therefore, its actual hidden capacity will be reduced; secondly, its secret image itself has a lower resolution.

[0007] Then, Chen Xianyi combined the mapping and generative carrier-free information hiding of images and proposed a new steganographic method based on natural image selection and StarGAN image generation network. This method divides the secret information into two parts, and then retrieves a suitable image from the database to represent the first part of the secret information during the image selection process; for the second part, an image mapping is established between it and the attribute features, and then the attribute features are modified according to the mapping relationship and the selected image, and the steganographic face image is generated using StarGAN. This method makes up for the shortcoming of Bi Xinliang's method that the steganographic capacity is much smaller than the number of labels in the training data set in practical applications, and selects labels without semantic contradictions from the same training data set as training labels. The actual steganographic capacity of each steganographic image exceeds 33 bits. However, its resolution is still low and it is relatively blurry when observed by the human eye. Summary of the invention

[0008] The primary purpose of the present invention is to overcome the problems existing in the prior art and provide a steganography method based on generative images. The present invention can improve the capacity, security and robustness of generative image steganography.

[0009] In order to achieve the above object, the present invention provides a steganographic method based on generative images, the method comprising the following steps: S1: Determine secret information, a shared key and a first feature group set, and encode the secret information to obtain a binary string; S2: acquiring a second feature group set according to the shared key and the first feature group set; S3: Obtain a first feature text set according to the second feature group set and the binary string; S4: Perform the following steps on each feature text in the first feature text set: S4.1: Selecting one feature text from the first feature text set as an input feature text, and obtaining a style manipulation intensity and a disentanglement threshold matching the input feature text from a parameter database; S4.2: Inputting the carrier image, the input feature text, the style manipulation strength and the disentanglement threshold into an image style editor to obtain a style image; S4.3: Filtering a feature group matching the input feature text from the second feature group set; S4.4: Input the style image and the feature group into an image semantic extractor, obtain the matching similarity between each feature text in the feature group and the style image, input each matching similarity into an activation function, and obtain the matching probability corresponding to each matching similarity; S4.5: Determine whether the feature text corresponding to the maximum matching probability satisfies both the first condition and the second condition. If so, use the style image as the carrier image for the next iteration; otherwise, increase the style manipulation intensity and repeat steps S4.2 to S4.5. The final encrypted image is obtained after the iteration ends. S5: Use the encrypted image to perform secret communication.

[0010] Furthermore, step S2 specifically includes: Inputting the shared key and the first feature group set into a mapping relationship generation function to obtain a filtered feature group set, wherein the mapping relationship generation function is used to determine the gender and age group of the face image to be generated according to the shared key, and then filtering out feature groups unique to other genders and other age groups in the first feature group set; Generating random factors according to the shared key that are equal to the number of feature groups in the screened feature group set; An algorithm is used to scramble the screened feature set according to the random factor to obtain a second feature set.

[0011] Furthermore, step S3 specifically includes: The second feature group set includes a plurality of feature groups, obtaining the number of feature texts included in a feature group in the second feature group set, and defining an empty feature text set; The steganographic capacity of the feature group is calculated according to the number of feature texts, and the specific calculation is determined by the following formula:

[0012] in, is the steganographic capacity, in units of ; is the number of feature texts; Extracting from the binary string a binary substring having the same number as the steganographic capacity; Binary-encode the feature texts contained in the feature group in sequence to obtain binary codes, match the corresponding feature texts in the feature group according to the binary substrings and the binary codes, and insert the feature texts into the empty feature text set; Determine whether the binary string is empty. If it is empty, obtain the first feature text set; if it is not empty, traverse the second feature group set, execute the above steps, and obtain the first feature text set.

[0013] Furthermore, the extracting of the same number of binary codes as the steganographic capacity from the binary string specifically includes: if the length of the current binary string is less than the steganographic capacity, filling 0 at the head of the current binary string until the length of the current binary string is equal to the steganographic capacity, and then extracting.

[0014] Furthermore, in step S4.2, the carrier image is obtained from the face dataset according to the shared key.

[0015] Furthermore, the image style editor described in step S4.2 is a StyleCLIP model.

[0016] Furthermore, the semantic extractor described in step S4.4 is a CLIP model.

[0017] Furthermore, the matching similarity calculation described in step S4.4 is determined by the following formula:

[0018] in, is the image vector, is a text vector.

[0019] Furthermore, in step S4.5, the first condition is whether the feature text is identical to the input feature text; and the second condition is whether the matching probability corresponding to the feature text is greater than a threshold.

[0020] Furthermore, step S5 specifically includes: The receiver inputs the shared key and the first feature group set into the mapping relationship generation function, and outputs a second feature group set; Inputting each feature group in the second feature group set and the encrypted image into the extractor in sequence to extract feature text; The extracted feature text is subjected to a reverse mapping operation according to the second feature group set to obtain secret information.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. Compared with the existing technology, the steganographic capacity is greatly improved, reaching 40 to 70 bits / image; 2. Image tags have no semantic conflicts and can generate real face images that conform to natural phenomena, thus improving the security of steganographic communication; 3. The image resolution can reach up to 1024×1024, which greatly improves the visual clarity. Frequent communication is unlikely to arouse the suspicion of attackers, further improving the security of steganographic communication. 4. The accuracy of information extraction of the decryption system depends largely on the resolution of the secret image and the semantics contained in the secret image. The secret image generated by the present invention has high clarity and no semantic conflict, so the generated secret image has high robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of a steganographic method based on a generative image according to an embodiment of the present invention; Figure 2 is a flow chart of step S4 of a steganographic method based on a generative image according to an embodiment of the present invention; Figure 3 is a model structure diagram of the image style editor StyleCLIP according to an embodiment of the present invention; Figure 4 is the collinear relationship between text and image in the CLIP space of the embodiment of the present invention; Figure 5 It is a flowchart of a steganographic method based on a generative image according to an embodiment of the present invention; Figure 6 It is a mapping relationship construction process between secret information and feature text in a steganographic method based on a generative image according to an embodiment of the present invention; Figure 7 This is the information hiding process of an embodiment of the present invention; Figure 8 This is the information extraction process of the embodiment of the present invention. DETAILED DESCRIPTION

[0023] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0024] Example 1 like Figure 1 and 2 As shown, a steganographic method based on a generative image in an embodiment of the present invention comprises the following steps: S1: Determine secret information, a shared key and a first feature group set, and encode the secret information to obtain a binary string; S2: acquiring a second feature group set according to the shared key and the first feature group set; S3: Obtain a first feature text set according to the second feature group set and the binary string; S4: Perform the following steps on each feature text in the first feature text set: S4.1: Selecting a feature text from the first feature text set as an input feature text, and obtaining a style manipulation intensity and a disentanglement threshold matching the input feature text from a parameter database; S4.2: Input the carrier image, input feature text, style manipulation strength and disentanglement threshold into the image style editor to obtain a style image; S4.3: Filtering a feature group matching the feature text from the second feature group set; S4.4: Input the style image and the feature group into the image semantic extractor, obtain the matching similarity between each feature text in the feature group and the style image, input each matching similarity into the activation function, and obtain the matching probability corresponding to each matching similarity; S4.5: Determine whether the feature text corresponding to the maximum matching probability satisfies both the first condition and the second condition. If so, use the style image as the carrier image for the next iteration; otherwise, increase the style manipulation intensity and repeat steps S4.2 to S4.5. The final encrypted image is obtained after the iteration ends. S5: Using encrypted images for secret communication.

[0025] Compared with the prior art, the steganographic capacity of this embodiment is greatly improved, reaching 40 to 70 bits / image; Image labels have no semantic conflicts, and real face images that conform to natural phenomena can be generated, thereby improving the security of steganographic communication; the image resolution can reach up to 1024×1024, and the visual clarity is greatly improved. Frequent communications are unlikely to arouse suspicion from attackers, further improving the security of steganographic communication; the accuracy of information extraction in the decryption system depends largely on the resolution of the secret image and the semantics contained in the secret image. The secret image generated by the present invention has high clarity and no semantic conflicts, so the generated secret image has high robustness.

[0026] Example 2 like Figures 1 to 8 As shown, a steganographic method based on a generative image in a preferred embodiment of the present invention comprises the following steps: S1: Determine secret information, a shared key and a first feature group set, and encode the secret information to obtain a binary string; In one embodiment, before the two parties communicate, a key shared by both parties is generated to construct a mapping relationship between secret information and feature text. To construct the mapping relationship between secret information and text, this embodiment proposes an initial feature group set: ,in Represented as the initial feature set The initial feature group set is a first feature group set, and the initial feature group includes feature groups such as "pupil color", "facial expression", "lady's hairstyle" and "beard shape"; feature group ,in Represented as feature group For example, the “hair color” feature group contains “blonde hair color”, “ginger hair color”, “black hair color” and other feature texts. It should be noted that the steganographic capacity of a feature group is determined by the number of feature texts it contains, that is, the number of feature texts contained in a feature group is , the steganalysis capacity of this feature group is .

[0027] Considering that the final generated encrypted image should not easily arouse the attacker's suspicion visually, feature groups and feature texts that conform to the characteristics of natural face images should be defined, and feature groups should be divided based on features such as "gender" and "age group". Taking gender features as an example, the feature groups unique to women include "women's hairstyle" and "lipstick color", and the feature groups unique to men include "men's hairstyle" and "beard shape", while feature groups such as "hair color", "pupil color" and "skin color" are common feature groups.

[0028] If you receive secret information , which is encoded into a binary string .

[0029] S2: acquiring a second feature group set according to the shared key and the first feature group set; In one embodiment, step S2 specifically includes: inputting the shared key and the first feature group set into a mapping relationship generation function to obtain a filtered feature group set, wherein the mapping relationship generation function is used to determine the gender and age group of the face image to be generated according to the shared key, and then filtering out feature groups unique to other genders and other age groups in the first feature group set; Generate random factors equal to the number of feature groups in the filtered feature group set according to the shared key; The algorithm is used to scramble the selected feature group set according to the random factor to obtain a second feature group set.

[0030] The mapping relationship generation function proposed in this embodiment is: , whose input is the shared key and the initial feature set , the function is based on Determine the gender and age group of the face image to be generated, and filter out the feature groups unique to other genders and other age groups in the initial feature group set; then Generate a random factor that matches the number of feature groups; use the Fisher–Yates algorithm to shuffle the order of the remaining feature groups in the initial feature group set according to the random factor to obtain the feature group set , This is the second feature group set. S3: Obtain a first feature text set according to the second feature group set and the binary string; In a specific embodiment, S3 specifically includes: The second feature group set includes multiple feature groups. The number of feature texts contained in a feature group in the second feature group set is obtained, and an empty feature text set is defined. ; The steganographic capacity of the feature group is calculated based on the number of feature texts. The specific calculation is determined by the following formula:

[0031] in, is the steganographic capacity, in units of ; is the number of feature texts; From binary string The same number of binary substrings as the steganographic capacity are intercepted from ; From the binary string The same number of binary codes as the stego capacity are intercepted, specifically including: if the length of the current binary string is less than the stego capacity, 0 is filled in the head of the current binary string until the length of the current binary string is equal to the stego capacity, and then interception is performed.

[0032] Binary encode the feature texts contained in the feature group in sequence to obtain the binary code, match the corresponding feature texts in the feature group according to the binary substring and the binary code, and insert the feature texts into the empty feature text set ; Determine whether the binary string is empty. If it is empty, obtain the first feature text set; if it is not empty, traverse the second feature group set, execute the above steps, and obtain the first feature text set.

[0033] S4: Perform the following steps on each feature text in the first feature text set: S4.1: Select a feature text from the first feature text set as the input feature text , obtaining the style manipulation strength and disentanglement threshold matching the input feature text from the parameter database; In one embodiment, the style control strength Control the number of channels affected by image style migration and increase the parameter It will enhance the effect of feature text on image transfer style, thereby improving the extraction rate of feature text by the extractor, and vice versa; disentanglement threshold Control the disentanglement degree of image style transfer. Increasing the threshold will reduce the influence of feature text on other features of the image in style transfer, and vice versa. In summary, the generator can set appropriate initial parameters for different feature texts. and , to improve the quality of the generated encrypted image and the extraction rate of the feature text, thereby improving the efficiency of the generated encrypted image. Through experimental results comparison, the corresponding appropriate parameters are debugged for all feature texts. and , a parameter database was constructed , style control intensity Disentanglement Threshold From the parameter database Select from.

[0034] S4.2: Carrier image, input feature text, style manipulation intensity Disentanglement Threshold Input into the image style editor to obtain the style image; In one embodiment, the image style editor in step S4.2 is a StyleCLIP model. Figure 3 This embodiment is based on the improvement of the Global Direction method proposed by StyleCLIP, and uses the powerful facial feature editing potential of the StyleGAN model to achieve multi-style control of facial images. The StyleCLIP model has the following steps: First, use an image encoder based on the StyleGAN structure to map the image to Then, any vector Through affine transformation Transform to Space, such as Figure 3 shown. Each direction in the space represents a specific image feature, and the key to image feature editing is to find the direction that is closely related to the feature text. For the style code, The corresponding generated image. Find the operation direction based on the feature text of the given target attribute ,so Images can be produced with this property introduced or amplified without significantly affecting other properties. The strength of the manipulation is determined by Control. Figure 4 As shown in , in the well-trained regions of the CLIP space, the directions on the text and image manifolds corresponding to the same semantic changes are roughly collinear and almost the same after normalization. Should be The direction of the CLIP space defined by the feature text represented by is consistent. According to the relationship between the above increments, we can use the collinearity between the text and image manifolds in the CLIP space to map a pair of text prompts based on prompt engineering to the CLIP space, and obtain , then roughly get Next, evaluate Each channel With a given direction in CLIP space Generate a style code By adding a negative value and a positive value, only the perturbation of each style code Channel. represents the CLIP space direction between the generated image pairs, and the channel The correlation with the target operation is estimated to be exist The average projection on is given by

[0035] When evaluating the relevance of each channel Afterwards, it ignores Below threshold Using a higher threshold will result in more disentanglement operations, but will reduce the visual effect of the operation. Since various high-level attributes (such as gender) involve the combination of several low-level attributes (such as hairstyle, eyebrows, and lips), lowering the threshold may be appropriate when multiple channels are correlated. Based on the above work, we get a To control the specific characteristics of the cover image style changes. The calculation formula is

[0036] get After that, Corresponding style codes for each operation direction in space implement , to generate the target style image, the obtained style image is .

[0037] It should be noted that the carrier image Obtained from the face dataset based on the shared key.

[0038] S4.3: Filtering a feature group matching the input feature text from the second feature group set; In one embodiment, from the second feature set Filter and feature text Matching feature groups ; S4.4: Input the style image and the feature group into the image semantic extractor, obtain the matching similarity between each feature text in the feature group and the style image, input each matching similarity into the activation function, and obtain the matching probability corresponding to each matching similarity; In one embodiment, the style image and feature groups Input image semantic extractor to obtain feature set Each feature text and style image The matching similarity of each match is input separately Activation function, obtains the matching probability corresponding to each matching similarity; It should be noted that the semantic extractor is a CLIP model. The matching similarity calculation is determined by the following formula:

[0039] in, is the image vector, is a text vector.

[0040] S4.5: Determine whether the feature text corresponding to the maximum matching probability satisfies both the first condition and the second condition. If so, use the style image as the carrier image for the next iteration; otherwise, increase the style manipulation intensity and repeat steps S4.2 to S4.5. The final encrypted image is obtained after the iteration ends. It should be noted that the first condition is the feature text With input feature text Are they the same? The second condition is the feature text Whether the corresponding matching probability is greater than the threshold.

[0041] S5: Using encrypted images for secret communication.

[0042] When receiving the encrypted image, first use the shared key and the initial feature set Input mapping relationship generation function Get feature group set , in order to establish the mapping relationship between secret information and feature text; then use the extractor to extract the feature text from the secret image; finally, map the extracted feature text back to the secret information according to the reverse operation of the mapping between secret information and feature text. The process of extracting feature text is as follows Figure 8 shown.

[0043] Example 3 This embodiment also provides a computer-readable storage medium, on which is stored a computer program for a steganographic method based on a generative image. When the computer program for a steganographic method based on a generative image is processed, the steps of the steganographic method based on a generative image are implemented.

[0044] In summary, the embodiments of the present invention provide a steganographic method based on generative images, which has a higher steganographic capacity than the prior art, reaching 40 to 70 bits / image; the image labels have no semantic conflicts, and can generate real face images that conform to natural phenomena, thereby improving the security of steganographic communication; the image resolution can reach up to 1024×1024, and the visual clarity is greatly improved. Frequent communications are also unlikely to arouse the suspicion of attackers, further improving the security of steganographic communication; the accuracy of information extraction of the decryption system part depends largely on the resolution of the secret image and the semantics contained in the secret image. The secret image generated by the present invention has high clarity and no semantic conflicts, so the generated secret image has high robustness.

[0045] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.

Claims

1. A steganographic method based on generative images, characterized in that: The method comprises the following steps: S1: Determine secret information, a shared key and a first feature group set, and encode the secret information to obtain a binary string; S2: acquiring a second feature group set according to the shared key and the first feature group set; S3: Obtain a first feature text set according to the second feature group set and the binary string; S4: Perform the following steps on each feature text in the first feature text set: S4.1: Selecting one feature text from the first feature text set as an input feature text, and obtaining a style manipulation intensity and a disentanglement threshold matching the input feature text from a parameter database; S4.2: Inputting the carrier image, the input feature text, the style manipulation strength and the disentanglement threshold into an image style editor to obtain a style image; S4.3: Filtering a feature group matching the input feature text from the second feature group set; S4.4: Input the style image and the feature group into an image semantic extractor, obtain the matching similarity between each feature text in the feature group and the style image, input each matching similarity into an activation function, and obtain the matching probability corresponding to each matching similarity; S4.5: Determine whether the feature text corresponding to the maximum matching probability satisfies both the first condition and the second condition. If so, use the style image as the carrier image for the next iteration; otherwise, increase the style manipulation intensity and repeat steps S4.2 to S4.

5. The final encrypted image is obtained after the iteration ends. S5: Use the encrypted image to perform secret communication.

2. According to claim 1, a steganographic method based on generative images is characterized in that: Step S2 specifically includes: Inputting the shared key and the first feature group set into a mapping relationship generation function to obtain a filtered feature group set, wherein the mapping relationship generation function is used to determine the gender and age group of the face image to be generated according to the shared key, and then filtering out feature groups unique to other genders and other age groups in the first feature group set; Generating random factors according to the shared key that are equal to the number of feature groups in the screened feature group set; An algorithm is used to scramble the screened feature set according to the random factor to obtain a second feature set.

3. According to the steganographic method based on generative image in claim 1, it is characterized in that: Step S3 specifically includes: The second feature group set includes a plurality of feature groups, obtaining the number of feature texts included in a feature group in the second feature group set, and defining an empty feature text set; The steganographic capacity of the feature group is calculated according to the number of feature texts, and the specific calculation is determined by the following formula: in, is the steganographic capacity, in units of ; is the number of feature texts; Extracting from the binary string a binary substring having the same number as the steganographic capacity; Binary-encode the feature texts contained in the feature group in sequence to obtain binary codes, match the corresponding feature texts in the feature group according to the binary substrings and the binary codes, and insert the feature texts into the empty feature text set; Determine whether the binary string is empty. If it is empty, obtain the first feature text set; if it is not empty, traverse the second feature group set, execute the above steps, and obtain the first feature text set.

4. A steganographic method based on generative images according to claim 3, characterized in that: The method of extracting the same number of binary codes as the steganographic capacity from the binary string specifically includes: if the length of the current binary string is less than the steganographic capacity, filling 0 at the head of the current binary string until the length of the current binary string is equal to the steganographic capacity, and then extracting the binary codes.

5. A steganographic method based on generative images according to claim 1, characterized in that: In step S4.2, the carrier image is obtained from the face dataset according to the shared key.

6. A steganographic method based on generative images according to claim 1, characterized in that: The image style editor described in step S4.2 is the StyleCLIP model.

7. A steganographic method based on generative images according to claim 1, characterized in that: The semantic extractor described in step S4.4 is the CLIP model.

8. A steganographic method based on generative images according to claim 7, characterized in that: The matching similarity calculation described in step S4.4 is determined by the following formula: in, is the image vector, is a text vector.

9. A steganographic method based on generative images according to claim 1, characterized in that: Step S4.5 The first condition is whether the feature text is the same as the input feature text; the second condition is whether the matching probability corresponding to the feature text is greater than a threshold.

10. A steganographic method based on generative images according to any one of claims 1 to 9, wherein The characteristic is that Step S5 specifically includes: The receiver inputs the shared key and the first feature group set into the mapping relationship generation function, and outputs a second feature group set; Inputting each feature group in the second feature group set and the encrypted image into the extractor in sequence to extract feature text; The extracted feature text is subjected to a reverse mapping operation according to the second feature group set to obtain secret information.

Images