Training device, training method, training program, data encoder, data encoding method, data encoding program, data decoder, data decoding method, and data decoding program

Abstract

The objective is to provide a steganography technique with high steganization resistance. [Solution] One aspect of the present disclosure relates to a training device comprising: a model processing unit that generates a cover image from an input image using an image generation model; a steganography processing unit that performs steganography encoding on the cover image and secret data to generate a stego image; a steganography processing unit that performs steganography on the stego image and obtains detection results; and a model training unit that trains the image generation model according to the loss value of the detection results calculated by a predetermined loss function.

Description

Technical Field

[0001] The present disclosure relates to a training device, a training method, a training program, a data encoder, a data encoding method, a data encoding program, a data decoder, a data decoding method, and a data decoding program.

Background Art

[0002] There is a technique called steganography as a means of embedding data in images, voices, videos, texts, etc. and secretly transmitting information. Also, there is a technique called steganalysis as a means of detecting steganography or decrypting data embedded by steganography. Typically, the images, voices, videos, texts, etc. here are called cover data, and the data embedded in the cover data can be called secret data.

[0003] In image steganography, secret data is embedded by changing the color information or luminance information of pixels in a cover image. In audio steganography, secret data is embedded by changing the amplitude or frequency of a cover audio. In video steganography, secret data is embedded in each frame or audio of a cover video. In text steganography, secret data is embedded in specific words, orders, blanks, line breaks, etc.

[0004] Steganography is used in various fields such as prevention of information leakage and secure transfer of data. On the other hand, there are also cases where it is used for illegal purposes such as viruses and malware. Steganography is characterized by not making it perceptible that information is hidden, unlike encryption. That is, unlike encryption, the existence of secret data cannot be recognized at first glance, so attacks from third parties can be avoided. Information protection technologies utilizing such characteristics of steganography have been proposed.

Prior Art Documents

Patent Documents

[0005] [Patent Document 1] Japanese Patent Publication No. 2010-93813 [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] Traditional steganography techniques require the creation of a cover image. If an image already available on the web is used as a cover image, there is a risk that an attacker could detect the presence of steganography through pixel differences, etc. Therefore, it is desirable to use an original image for the cover image.

[0007] On the other hand, when building a system for sending and receiving confidential data based on steganography, it is necessary to prepare original images periodically, which becomes a significant burden.

[0008] In light of the above issues, the objective of this disclosure is to provide a steganography technology with high steganization resistance. [Means for solving the problem]

[0009] One aspect of the present disclosure relates to a training device comprising: a model processing unit that generates a cover image from an input image using an image generation model; a steganography processing unit that performs steganography encoding on the cover image and secret data to generate a stego image; a steganography processing unit that performs steganography on the stego image and obtains detection results; and a model training unit that trains the image generation model according to the loss value of the detection results calculated by a predetermined loss function. [Effects of the Invention]

[0010] According to this disclosure, it is possible to provide a steganography technology with high steganization resistance. [Brief explanation of the drawing]

[0011] [Figure 1] Figure 1 is a schematic diagram showing a steganography process according to one embodiment of the present disclosure. [Figure 2] Figure 2 is a schematic diagram showing the retraining process of an image generation model according to one embodiment of the present disclosure. [Figure 3] Figure 3 is a block diagram showing the hardware configuration of a training device, data encoder, and data decoder according to one embodiment of the present disclosure. [Figure 4] Figure 4 is a block diagram showing the functional configuration of a training device according to one embodiment of the present disclosure. [Figure 5] Figure 5 shows the architecture of a Stable Diffusion according to one embodiment of the present disclosure. [Figure 6] Figure 6 is a flowchart showing the training process according to one embodiment of the present disclosure. [Figure 7] Figure 7 is a block diagram showing the functional configuration of a data encoder according to one embodiment of the present disclosure. [Figure 8] Figure 8 is a block diagram showing the functional configuration of a data decoder according to one embodiment of the present disclosure. [Figure 9] Figure 9 is a sequence diagram showing a steganography process according to one embodiment of the present disclosure. [Modes for carrying out the invention]

[0012] Embodiments of this disclosure will be described below with reference to the drawings.

[0013] In the following embodiment, a training device for training an image generation model that generates cover images used in steganography, and a data encoder and data decoder for exchanging data using steganography with the trained image generation model are disclosed.

[0014] [Summary of this disclosure] In the steganography process according to the following embodiments, as shown in FIG. 1, when a cover image generated by an image generation model 20 and secret data to be embedded in the cover image are given, a data encoder 200 encodes the secret data into the cover image by any known steganography method and generates a stego image in which the secret data is embedded. Then, the data encoder 200 transmits the generated stego image to a data decoder 300 at the destination of the secret data. When receiving the stego image from the data encoder 200, the data decoder 300 decodes the received stego image and extracts the secret data from the stego image.

[0015] The image generation model 20 according to the illustrated embodiment can be realized as any known machine learning model such as Stable Diffusion. In the embodiments described later, the image generation model 20 is retrained from any known machine learning model such as Stable Diffusion so as to be suitable for generating a cover image used for image steganography to enhance steganalysis resistance.

[0016] Specifically, as shown in FIG. 2, a training device 100 inputs an image extracted from a retraining image dataset into the image generation model 20 and obtains a cover image from the image generation model 20. Then, the training device 100 generates a stego image from the cover image and the secret data by any known steganography method, and based on a loss value calculated by a loss function from the error between the input image to the image generation model 20 and the cover image and the detection result by steganalysis for the stego image generated from the cover image, adjusts the parameters of the image generation model 20. The training device 100 executes parameter adjustment for each image in the image dataset, and when the parameter adjustment is completed for all the input images prepared for retraining, finally provides the obtained image generation model 20 to the data encoder 200.

[0017] Each term used in the following examples is as follows. A cover image is an image for embedding secret data. Secret data is data to be embedded, and is not limited to images as illustrated, and may be of any data type. A stego image is an image in which secret data has been embedded by steganography. Encoding means embedding data, and decoding means extracting the embedded data.

[0018] Here, the training device 100, the data encoder 200, and the data decoder 300 may have a hardware configuration as shown in, for example, FIG. 3. That is, the training device 100, the data encoder 200, and the data decoder 300 have an interface device 101, a storage device 102, a memory device 103, a processor 104, a user interface (UI) device 105, and a communication device 106 that are interconnected via a bus B.

[0019] Programs and / or data for realizing various functions and processes in the training device 100, the data encoder 200, and the data decoder 300 are downloaded from an external device and / or network to the storage device 102 or the memory device 103 via the interface device 101.

[0020] The storage device 102 is realized by a non-volatile memory or the like, and stores installed or downloaded programs or data (for example, files, etc.).

[0021] The memory device 103 is realized by a random access memory, a static memory, etc., and when a program or instruction is started, reads and stores the program, instruction, data, etc. from the storage device 102. The storage device 102, the memory device 103, and the removable storage medium may be collectively referred to as a non-transitory storage medium.

[0022] The processor 104 may be implemented by one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), processing circuits, etc., which may consist of one or more processor cores, and executes various functions and processes of the training device 100, data encoder 200, and data decoder 300 according to data such as programs, instructions, and parameters necessary to execute the programs or instructions stored in the memory device 103.

[0023] The user interface (UI) device 105 may consist of input devices such as a keyboard, mouse, camera, and microphone, output devices such as a display, speaker, headset, and printer, and input / output devices such as a touch panel, and realizes an interface between the user and the training device 100, data encoder 200, and data decoder 300. For example, the user may operate the training device 100, data encoder 200, and data decoder 300 by operating a GUI (Graphical User Interface) displayed on a display or touch panel with a keyboard, mouse, etc.

[0024] The communication device 106 is implemented by various communication circuits that perform wired and / or wireless communication processing with external devices, the Internet, LAN (Local Area Network), cellular networks, and other communication networks.

[0025] However, the hardware configuration described above is merely an example, and the training device 100, data encoder 200, and data decoder 300 according to this disclosure may be implemented by any other suitable hardware configuration.

[0026] [Training device] A training device 100 according to one embodiment of the present disclosure will be described. Figure 4 is a block diagram showing the functional configuration of the training device 100 according to one embodiment of the present disclosure. As shown in Figure 4, the training device 100 has a model processing unit 110, a steganography processing unit 120, a steganization processing unit 130, and a model training unit 140. For example, one or more of the functional units of the model processing unit 110, steganography processing unit 120, steganization processing unit 130, and model training unit 140 may be realized by one or more processors 104 executing one or more programs or instructions.

[0027] The model processing unit 110 generates a cover image from the input image using the image generation model 20. For example, the image generation model 20 may be any known deep learning-based image generation model such as Stable Diffusion, GAN, or VAE, and is retrained by the training device 100 to be suitable for generating cover images for steganography.

[0028] Specifically, an image dataset is prepared for retraining the image generation model 20. The model processing unit 110 inputs each image from the image dataset into the image generation model 20 to be retrained, and obtains a cover image from the image generation model 20. The model processing unit 110 provides the cover image generated by the image generation model 20 to the encoding processing unit 120.

[0029] The steganography processing unit 120 performs steganographic encoding on the cover image and the secret data to generate a stego image. Here, the secret data may be any data that you want to send secretly to the recipient by embedding it in the cover image. For example, in a distribution system for books, videos, music, etc., the secret data may be content data such as book page images, video frames, or music tracks. Alternatively, the secret data may be encrypted, and the encrypted secret data may be embedded in the cover image.

[0030] In this embodiment, the steganography processing unit 120 may perform steganographic encoding on the cover image and secret data according to any known image steganography method such as LSB (Least Significant Bit) steganography, DCT (Discrete Cosine Transform) steganography, DWT (Discrete Wavelet Transform) steganography, SteganoGAN (Generative Adversarial Network), or Cross.

[0031] Here, LSB steganography, DCT steganography, and DWT steganography are rule-based steganography methods, while SteganoGAN and Cross are machine learning-based steganography methods. Rule-based steganography methods embed sensitive data based on specific rules or features, making them highly susceptible to detection through statistical analysis, feature extraction, and machine learning-based steganization. Furthermore, if the algorithm used as a decoding tool is leaked, the sensitive data can be decrypted. On the other hand, machine learning-based steganography methods require machine learning models not only in the encoding environment but also in the decoding environment, posing implementation difficulties such as the need for high-end GPUs in the decoding environment, such as user terminals. Moreover, if the machine learning model used as a decoding tool is leaked, the sensitive data can be decrypted.

[0032] For example, in LSB steganography encoding, the secret data is first converted into a bit sequence of "0"s and "1"s. Meanwhile, the bits of each pixel in each RGB color space of the cover image are replaced with the bit sequence of the secret data, starting from the least significant bit. Here, the concealment of the secret data is enhanced by embedding it in areas that are difficult for the human eye to perceive or in areas that are less affected by image changes. The embedding location information can be stored in another file, or it can be embedded in the cover image, just like the secret data.

[0033] On the other hand, in the LSB steganography decoding process, embedding location information is obtained, and the bit sequence of each pixel in each RGB color space of the stego image is extracted from the embedding location indicated by the obtained embedding location information. The secret data is then restored from the extracted bit sequence.

[0034] The steganography processing unit 120 performs an encoding process using one of the steganography methods on the cover image generated by the image generation model 20 and the secret data to be embedded, and obtains a stego image in which the secret data is encoded in the cover image. The steganography processing unit 120 then provides the obtained stego image to the steganization processing unit 130.

[0035] The steganography processing unit 130 performs steganography on the stego image and obtains the detection results. Specifically, the steganography processing unit 130 performs steganography on the stego image according to any known image steganography method. Such steganography methods include targeted steganography, which is specialized for a particular steganography method, such as detecting LSB steganography, and universal steganography, which detects any steganography method. In this embodiment, the objective is to improve resistance to targeted steganography and universal steganography aimed at detecting image steganography.

[0036] The steganography processing unit 130 provides the model training unit 140 with the detection results of steganography performed on the stego image generated by the steganography processing unit 120. For example, the detection results of steganography may be a value of 0.0 to 1.0 indicating the likelihood or probability of steganography, a value of 0.0 or greater indicating the amount of steganographic embedding, or a true / false truth value indicating whether or not the data has been processed with steganography.

[0037] The model training unit 140 trains the image generation model 20 according to the loss value of the detection result calculated by a predetermined loss function. For example, the image generation model 20 may be implemented by Stable Diffusion, and as an example, it may consist of an architecture as shown in Figure 5. According to the illustrated architecture, the image generation model 20 consists of a VAE (Variational Auto-Encoder) encoder 21 that converts the input image into latent variables, a text encoder 22 that converts the input prompt into a vector representation, a U-net type diffusion model 23 that repeatedly adds noise to the image (latent variables) to generate an image of only noise, and then gradually removes the noise from the noise image to restore the original image (latent variables), and a VAE decoder 24 that converts the latent variables into a final image. That is, when the image generation model 20 receives an input image and a prompt, it generates a cover image. If no input image or input prompt is specified, the image generation model 20 generates an image using random noise as a seed.

[0038] In one embodiment, the model training unit 140 may train the VAE decoder 24 according to the loss value. In other words, the model training unit 140 does not retrain the VAE encoder 21, the text encoder 22, and the diffusion model 23. This makes the training process more efficient by retraining only the downstream VAE decoder 24 without retraining the entire image generation model 20.

[0039] Here, the predetermined loss function may determine the loss value based on the image error between the input image and the cover image and the detection accuracy of the steganization. For example, the loss function Loss is: Loss = Loss(A) + Loss(B) * α This may also be the case. Here, Loss(A) is the original loss function of the image generation model 20 being trained, and may be based on the error between the input image to the image generation model 20 and the cover image output from the image generation model 20. Also, Loss(B) is a loss function based on the detection result of steganization, and may be based on the detection accuracy of steganization. Furthermore, α is a weight coefficient between 0 and 1.

[0040] Specifically, Loss(A) may be any of the following: Loss(A) = MSE*α + LPIPS*β + KL*γ Here, MSE is the mean squared error between the predicted value (cover image) and the true value (input image), LPIPS is the human-perceived inter-image distance between the predicted value and the true value, and KL is the KL (Kullback-Leibier) divergence, which is the error between the distribution of the latent variable (output of the VAE encoder 21) and the standard normal distribution. Also, α, β, and gamma waves are weighting coefficients between 0 and 1.

[0041] On the other hand, Loss(B) may be any of the following: Loss(B) = BCE Here, BCE is the binary cross-entropy, which is the error between the predicted value (likelihood under steganization) and the true value (not under steganization, likelihood = 0) in a binary classification.

[0042] The model training unit 140 retrains the image generation model 20 described above until predetermined termination conditions are met. Here, the predetermined termination conditions may be until the retraining process described above is executed a predetermined number of times, or until the loss value of the image generation model 20 falls below a predetermined threshold.

[0043] According to the training device 100 described above, an image generation model 20 capable of generating cover images with high steganization tolerance can be obtained.

[0044] [Training Process] Next, a training process according to one embodiment of the present disclosure will be described. This training process can be implemented by the training device 100, more specifically by the processor 104 of the training device 100 executing a program stored in the memory device 103. Figure 6 is a flowchart of the training process according to one embodiment of the present disclosure.

[0045] As shown in Figure 6, in step S101, the training device 100 generates a cover image from the input image using the image generation model 20. Specifically, the training device 100 extracts images from the image dataset and inputs the extracted images into the image generation model 20. The training device 100 then obtains the cover image from the image generation model 20.

[0046] In step S102, the training device 100 performs steganographic encoding on the cover image and the secret data to generate a Stego image. Specifically, the training device 100 performs any known steganographic encoding on the cover image generated by the image generation model 20 in step S101 and the secret data to be embedded in the cover image to obtain a Stego image in which the secret data is embedded in the cover image.

[0047] In step S103, the training device 100 performs steganography on the stego image and obtains the detection result. Specifically, the training device 100 performs any known steganography on the stego image in which secret data is embedded in the cover image, and obtains the likelihood of the steganography regarding whether steganography has been performed on the stego image as the detection result.

[0048] In step S104, the training device 100 trains the image generation model 20 according to the loss value of the detection result calculated by a predetermined loss function. Specifically, the training device 100 calculates the loss value of the loss function based on the image error between the input image and the cover image and the detection accuracy of the steganization, and trains the image generation model 20 according to the calculated loss value.

[0049] In step S105, the training device 100 determines whether the termination conditions have been met. For example, the termination conditions may be that steps S101 to S104 have been performed for all images in the image dataset prepared for retraining the image generation model 20, or that the loss value has converged to or below a predetermined threshold. If the termination conditions are met (S105: YES), the training process is terminated. On the other hand, if the termination conditions are not met (S105: NO), the training device 100 returns to step S101 and generates a cover image for the next image in the image dataset.

[0050] According to the training process described above, an image generation model 20 capable of generating cover images with high steganization tolerance can be obtained.

[0051] [Data Encoder] Next, a data encoder 200 according to one embodiment of the present disclosure will be described. The data encoder 200 generates a stego image from a cover image generated by an image generation model 20 retrained by a training device 100 and secret data to be embedded in the cover image by steganographic encoding, and transmits the stego image to, for example, a destination for the secret data.

[0052] Figure 7 is a block diagram showing the functional configuration of a data encoder 200 according to one embodiment of the present disclosure. As shown in Figure 7, the data encoder 200 has a data acquisition unit 210 and an encoding processing unit 220. For example, one or more functional units of the data acquisition unit 210 and the encoding processing unit 220 may be realized by one or more processors 104 executing one or more programs or instructions.

[0053] The data acquisition unit 210 acquires the cover image generated by the image generation model 20 and the secret data. The image generation model 20 can be implemented, for example, as a machine learning model retrained from Stable Diffusion to be suitable for generating steganography cover images by the training device 100 described above.

[0054] Such an image generation model 20 can be retrained by performing the following: generating a cover image from input images extracted from an image dataset; performing steganographic encoding on the cover image and secret data to generate a stego image; performing steganization on the stego image to obtain detection results; and training the image generation model 20 according to the loss value of the detection results calculated by a predetermined loss function.

[0055] For example, the image generation model 20 may generate a cover image based on related data concerning the secret data input to the Stable Diffusion text encoder 22. Such related data may include tag information and image information of the secret data. When generating the cover image, the tag information and image information of the secret data may be used to generate the image using text2image or image2image.

[0056] The encoding processing unit 220 performs steganographic encoding on the cover image and the secret data to generate a stego image. Specifically, the encoding processing unit 220 generates the stego image according to any known steganographic technique, the same one used by the training device 100.

[0057] Here, the encoding processing unit 220 may encrypt the secret data, perform steganography encoding on the cover image and the encrypted secret data, and generate a stego image. In this case, the encoding processing unit 220 needs to provide the stego image data decoder 300 with a decryption key corresponding to the encryption key used to encrypt the secret data, by any method of provision.

[0058] The encoding process 220 sends the generated Stego image to the destination of the secret data.

[0059] According to the data encoder 200 described above, a Stego image with embedded confidential data to be transmitted can be generated using a cover image that has high steganization resistance.

[0060] [Data Decoder] Next, a data decoder 300 according to one embodiment of the present disclosure will be described. The data decoder 300 obtains a stego image generated by steganographic encoding from a cover image generated by an image generation model 20 retrained by a training device 100 and secret data to be embedded in the cover image from the data encoder 200, and performs steganographic decoding on the stego image to obtain the secret data.

[0061] Figure 8 is a block diagram showing the functional configuration of a data decoder 300 according to one embodiment of the present disclosure. As shown in Figure 8, the data decoder 300 has an image acquisition unit 310 and a decoding processing unit 320. For example, one or more functional units of the image acquisition unit 310 and the decoding processing unit 320 may be realized by one or more processors 104 executing one or more programs or instructions.

[0062] The image acquisition unit 310 acquires a Stego image generated by performing steganographic encoding on the cover image generated by the image generation model 20 and the secret data. Specifically, the image acquisition unit 310 acquires a Stego image generated by the data encoder 200 described above.

[0063] The decoding processing unit 320 performs a decoding process on the Stego image and obtains secret data from the Stego image. Specifically, the secret data encoded in the Stego image may be content data. For example, the content data may be book data, video data, music data, etc.

[0064] In this case, the decoding processing unit 320 may only display the content data obtained through the decoding process without storing it. Specifically, the decoder processing unit 320 may not store the decoded content data in the data decoder 300, but only allow the user to view it on the viewer. This can prevent the content data from being duplicated or otherwise copied.

[0065] Furthermore, if the secret data is encrypted, the decoding processing unit 320 may decrypt the encrypted secret data extracted from the Stego image using a decryption key obtained from the data encoder 200 by any means.

[0066] According to the data decoder 300 described above, confidential data can be obtained from a Stego image generated from a cover image with high steganization resistance.

[0067] [Steganography processing] Next, a steganography process according to one embodiment of the present disclosure will be described. This steganography process can be realized by a data encoder 200 and a data decoder 300, and more specifically, by a processor 104 of the data encoder 200 and data decoder 300 executing a program stored in a memory device 103. Figure 9 is a sequence diagram showing a steganography process according to one embodiment of the present disclosure.

[0068] As shown in Figure 9, in step S201, the data encoder 200 acquires the cover image and the secret data. The cover image is generated by the image generation model 20, which has been retrained from Stable Diffusion or the like by the training device 100 described above, and the secret data is the data to be embedded in the cover image.

[0069] In step S202, the data encoder 200 performs steganographic encoding on the cover image and the secret data to generate a stego image in which the secret data is embedded in the cover image.

[0070] In step S203, the data encoder 200 transmits the Stego image to the data decoder 300.

[0071] In step S204, the data decoder 300 performs steganography decoding on the stego image.

[0072] In step S205, the data decoder 300 obtains secret data from the Stego image.

[0073] According to the steganography process described above, a cover image with high steganization resistance can be used to generate a stego image in which the confidential data to be transmitted is embedded, and the confidential data can be retrieved from the generated stego image.

[0074] Although embodiments of this disclosure have been described in detail above, this disclosure is not limited to the specific embodiments described above, and various modifications and changes are possible within the scope of the gist of this disclosure as described in the claims. [Explanation of Symbols]

[0075] 20 Image Generation Models 100 training equipment 110 Model Processing Unit 120 Steganography Processing Unit 130 Steganization Processing Unit 140 Model Training Department 200 data encoders 210 Data acquisition unit 220 Encoding Processing Unit 300 data decoders 310 Image acquisition unit 320 Decode Processing Unit

Claims

1. A model processing unit that generates a cover image from an input image using an image generation model, A steganography processing unit performs steganography encoding on the cover image and secret data to generate a stego image, A steganization processing unit performs steganization on the aforementioned stego image and obtains the detection result, A model training unit that trains the image generation model according to the loss value of the detection result calculated by a predetermined loss function, A training device having the following features.

2. The training apparatus according to claim 1, wherein the predetermined loss function determines the loss value based on the image error between the input image and the cover image and the detection accuracy of the steganization.

3. The training apparatus according to claim 1, wherein the image generation model is implemented by Stable Diffusion.

4. The training apparatus according to claim 3, wherein the model training unit trains the VAE (Variational Auto-Encoder) decoder of the Stable Diffusion according to the loss value.

5. Using an image generation model, generate a cover image from an input image, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, Perform steganization on the aforementioned stego image and obtain the detection results. The image generation model is trained according to the loss value of the detection result calculated by a predetermined loss function, A training method in which one or more computers perform the task.

6. Using an image generation model, generate a cover image from an input image, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, Perform steganization on the aforementioned stego image and obtain the detection results. The image generation model is trained according to the loss value of the detection result calculated by a predetermined loss function, A training program that causes one or more computers to execute a command.

7. A data acquisition unit that acquires a cover image generated by an image generation model and confidential data, An encoding processing unit that performs steganographic encoding on the cover image and secret data to generate a stego image, It has, The aforementioned image generation model is The process involves generating a cover image from an input image, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, Perform steganization on the aforementioned stego image and obtain the detection results. The image generation model is trained according to the loss value of the detection result calculated by a predetermined loss function, A data encoder that is trained by performing [a certain operation].

8. The image generation model is implemented by Stable Diffusion, as described in claim 7, for the data encoder.

9. The data encoder according to claim 8, wherein the image generation model generates the cover image based on relevant data relating to the confidential data input to the text encoder of the Stable Diffusion.

10. The data encoder according to claim 7, wherein the encoding processing unit encrypts the secret data, performs steganography encoding on the cover image and the encrypted secret data, and generates the stego image.

11. The process involves obtaining the cover image generated by the image generation model and the confidential data, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, It has, The aforementioned image generation model is The process involves generating a cover image from an input image, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, Perform steganization on the aforementioned stego image and obtain the detection results. The image generation model is trained according to the loss value of the detection result calculated by a predetermined loss function, Trained by performing A data encoding method performed by one or more computers.

12. The process involves obtaining the cover image generated by the image generation model and the confidential data, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, It has, The aforementioned image generation model is The process involves generating a cover image from an input image, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, Perform steganization on the aforementioned stego image and obtain the detection results. The image generation model is trained according to the loss value of the detection result calculated by a predetermined loss function, Trained by performing A data encoding program that runs on one or more computers.

13. An image acquisition unit that acquires a Stego image generated by performing steganographic encoding on a cover image generated by an image generation model and secret data, A decoding processing unit that performs a decoding process on the Stego image and obtains the secret data from the Stego image, It has, The aforementioned image generation model is The process involves generating a cover image from an input image, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, Perform steganization on the aforementioned stego image and obtain the detection results. The image generation model is trained according to the loss value of the detection result calculated by a predetermined loss function, A data decoder that is trained by performing the following.

14. The secret data encoded in the aforementioned Stego image is content data, The data decoder according to claim 13, wherein the decoding processing unit displays the content data obtained by the decoding process without storing it.

15. Obtaining a Stego image generated by performing steganographic encoding on a cover image generated by an image generation model and secret data, The process involves performing a decoding operation on the Stego image and obtaining the secret data from the Stego image, It has, The aforementioned image generation model is The process involves generating a cover image from an input image, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, Perform steganization on the aforementioned stego image and obtain the detection results. The image generation model is trained according to the loss value of the detection result calculated by a predetermined loss function, Trained by performing A data decoding method performed by one or more computers.

16. Obtaining a Stego image generated by performing steganographic encoding on a cover image generated by an image generation model and secret data, The process involves performing a decoding operation on the Stego image and obtaining the secret data from the Stego image, It has, The aforementioned image generation model is The process involves generating a cover image from an input image, The process involves performing steganographic encoding on the cover image and the secret data to generate a stego image, Perform steganization on the aforementioned stego image and obtain the detection results. The image generation model is trained according to the loss value of the detection result calculated by a predetermined loss function, Trained by performing A data decoding program that runs on one or more computers.

Images