Carrier-free image steganography based on diffusion model and human body posture estimation

Abstract

The invention relates to the technical field of carrier-free image steganography, and discloses a carrier-free image steganography based on a diffusion model and human body posture estimation, which comprises the following steps of: S1, converting secret information into a binary sequence according to an ASCII (American Standard Code for Information Interchange) code, and taking the binary sequence as a secret sequence to be encrypted; s2, extracting a large number of human whole body images from the network, according to a pre-trained human body posture estimation model, extracting posture key point information of a target image, key point arrangement features such as an angle between an arm and a trunk and the like, grouping a series of images, and according to a binary bit number set by each image, dividing the images into 2n groups; according to the carrier-free image steganography based on the diffusion model and the human body posture estimation, the encryption information capacity of a single picture is improved, the similarity of theme contents among multiple pictures is kept, the confidentiality and robustness are enhanced, the flexibility and expandability are improved, and the information transmission and storage efficiency is optimized; and efficient, safe and hidden information hiding and transmission are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of carrier - free image steganography, and specifically to a carrier - free image steganography based on a diffusion model and human pose estimation. Background Art

[0002] Carrier - free information steganography is a technology that does not rely on traditional carriers (such as images, audio, videos, etc.) to hide secret information during the information transmission process. It realizes the hiding and transmission of information by constructing the relationship between the secret information and specific information. This technology originated from the imitation of the "mimicry" phenomenon in the animal and plant kingdoms, aiming to improve the confidentiality and security of information transmission.

[0003] In the existing image steganography using diffusion models, semantics and labels are mostly used as public and private keys to guide image generation and encryption, aiming to use the high degree of correspondence between the content generated by the diffusion model and the labels and semantics to achieve the transmission of secret images through the asymmetric encryption of public and private keys.

[0004] However, due to technical limitations, the quality of image generation is relatively low, specifically reflected in the insufficient peak signal - to - noise ratio. As a result, the current technical level of the diffusion model cannot embed binary hash sequences or text like traditional models such as GAN (Generative Adversarial Network) to achieve the transmission of secret information. It can only transmit secret images, with extremely small encryption capacity and is not easy to achieve large - scale information transmission. At the same time, due to the label requirements of the generated images, the semantics of the public and private keys are relatively similar, and potential attackers can guess the semantics of the private key from the public key, resulting in insufficient security. Therefore, a carrier - free image steganography based on a diffusion model and human pose estimation is proposed to solve the above problems. Summary of the Invention

[0005] (1) Technical Problems to be Solved

[0006] Aiming at the deficiencies of the existing technology, the present invention provides a carrier - free image steganography based on a diffusion model and human pose estimation, which has the advantages of being able to generate images of different people with different backgrounds but similar postures, and improved concealment. It solves the problems of the existing technology that due to technical limitations, the quality of image generation is relatively low, specifically reflected in the insufficient peak signal - to - noise ratio, resulting in the current technical level of the diffusion model being unable to embed binary hash sequences or text like traditional models such as GAN (Generative Adversarial Network) to achieve the transmission of secret information, can only transmit secret images, with extremely small encryption capacity and is not easy to achieve large - scale information transmission. At the same time, due to the label requirements of the generated images, the semantics of the public and private keys are relatively similar, and potential attackers can guess the semantics of the private key from the public key, resulting in insufficient security.

[0007] (2) Technical Solutions

[0008] To achieve the above object, the present invention provides the following technical solutions: A carrier-free image steganography based on a diffusion model and human pose estimation, comprising the following steps:

[0009] S1. Convert the secret information into a binary sequence according to ASCII code, and wait to be encrypted as a secret sequence;

[0010] S2. Extract a large number of full-body human images from the network. According to the pre-trained human pose estimation model, extract the pose key point information of the target image, arrange the features of the key points, such as the angle between the arm and the torso, etc., to group the series of images, and divide them into 2^n groups according to the number of binary digits set for each image;

[0011] S3. Construct a mapping for each group with a binary sequence, and then take out the corresponding model from the corresponding group according to the binary sequence of the secret information;

[0012] S4. Input the pose key point information as a condition into the diffusion model to guide the model to generate an image consistent with the target pose;

[0013] S5. Appropriately reduce the corresponding image to achieve the splicing of multiple images, so that a large number of images with a long sequence are compressed into an acceptable number of images;

[0014] S6. Transmit the secret image to the receiver, and the receiver extracts the secret information according to the inverse process.

[0015] Preferably, the secret information in S1 includes at least one of text, numbers or symbols, and further includes encrypting or compressing the secret information to improve the security of the information or reduce the storage space of the information.

[0016] Preferably, the human pose estimation model in S2 is a deep learning-based model, including but not limited to OpenPose, AlphaPose or HRNet, and the pose key point information includes but not limited to key points such as head, shoulder, elbow, wrist, hip, knee, ankle, etc.

[0017] Preferably, the method of grouping the series of images in S2 includes extracting the feature of the pose key point information of each image, and performing clustering analysis according to the extracted features, and dividing the images with similar poses into the same group.

[0018] Preferably, the mapping relationship in S3 is a one-to-one correspondence relationship, and further includes encrypting the mapping relationship to prevent unauthorized access or tampering.

[0019] Preferably, the diffusion model in S4 is a generative model based on deep learning, including but not limited to DDPM or DDIM, and the conditional input includes at least one of pose key point information, image style, and lighting conditions.

[0020] Preferably, the reduction method in S5 includes but not limited to image cropping, image scaling, or image compression, and further includes encrypting or watermarking the reduced image to enhance the security of the image.

[0021] Preferably, the inverse process in S6 includes extracting pose key point information from the received image, grouping the image according to the extracted key point information, restoring the corresponding binary sequence according to the grouping result, and finally converting the binary sequence into the original secret information.

[0022] Preferably, the error detection and correction processing includes but not limited to CRC cyclic redundancy check or ECC error correction code technology.

[0023] (III) Beneficial effects

[0024] Compared with the prior art, the present invention provides a carrierless image steganography based on a diffusion model and human pose estimation, having the following beneficial effects:

[0025] 1. For the carrierless image steganography based on a diffusion model and human pose estimation, through the technology based on a diffusion model and human pose estimation, the steganography of the present invention can embed more secret information in a single picture. The diffusion model can generate high-quality and diverse images, while human pose estimation provides rich pose key point information, which can be used as an encryption carrier. Compared with traditional steganography, the present invention can significantly improve the encrypted information capacity of a single picture without significantly increasing the image complexity.

[0026] 2. For the carrierless image steganography based on a diffusion model and human pose estimation, grouping the images through human pose estimation makes multiple pictures basically consistent in the theme content, with only differences in the human actions. This design not only improves the visual consistency of the images but also enhances the concealment of the steganography. Since multiple pictures are visually similar, it is difficult for attackers to discover the hidden secret information through simple visual analysis, thereby improving the confidentiality of the system.

[0027] 3. The carrierless image steganography based on the diffusion model and human pose estimation combines the advantages of the diffusion model and human pose estimation to generate images that not only have a high degree of authenticity but also can resist common image processing attacks such as compression, cropping, and rotation. The images generated by the diffusion model have a high degree of diversity and complexity in details, making it difficult to detect and extract the hidden secret information. In addition, the pose key point information provided by human pose estimation is used as an encryption condition, enhancing the robustness of the system. Even if the image undergoes a certain degree of processing, the secret information can still be accurately extracted.

[0028] 4. The carrierless image steganography based on the diffusion model and human pose estimation can adapt to different application scenarios and requirements by adjusting the parameters of the diffusion model and the grouping strategy of human pose estimation.

[0029] 5. The carrierless image steganography based on the diffusion model and human pose estimation can significantly reduce the information transmission and storage costs by appropriately shrinking the images and stitching multiple images together, enabling the compression of a large number of images in a long sequence into an acceptable number of images. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic flow chart of a carrierless image steganography based on the diffusion model and human pose estimation of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0032] Please refer to Figure 1 , a carrierless image steganography based on the diffusion model and human pose estimation, including the following steps:

[0033] S1. Convert the secret information into a binary sequence according to ASCII code and wait to be encrypted as a secret sequence.

[0034] S2. Extract a large number of full-body human images from the network. According to the pre-trained human pose estimation model, extract the pose key point information of the target images, arrange the features of the key points, such as the angle between the arm and the torso, etc., to group the series of images, and divide them into 2^n groups according to the set number of binary digits for each image.

[0035] S3. Build a mapping for each group with a binary sequence, and then extract the corresponding model from the corresponding group according to the binary sequence of the secret information;

[0036] S4. Use the pose key point information as a condition to input into the diffusion model, and guide the model to generate an image consistent with the target pose;

[0037] S5. Appropriately reduce the corresponding image to achieve the splicing of multiple images, so that a large number of images in a long sequence are compressed into an acceptable number of images;

[0038] S6. Transmit the secret image to the receiver, and the receiver extracts the secret information according to the inverse process.

[0039] Specifically, the secret information in S1 includes at least one of text, numbers, or symbols, and further includes encrypting or compressing the secret information to improve the security of the information or reduce the storage space of the information.

[0040] Furthermore, the encryption process includes but is not limited to AES (Advanced Encryption Standard), RSA (asymmetric encryption algorithm), or DES (Data Encryption Standard), and the compression process includes but is not limited to ZIP, RAR, or LZ77 algorithm. In addition, before the secret information is converted into a binary sequence, it can be segmented, and the length of each segment can be flexibly adjusted according to actual needs to facilitate subsequent grouping and mapping operations.

[0041] Specifically, the human pose estimation model in S2 is a deep learning-based model, including but not limited to OpenPose, AlphaPose, or HRNet, and the pose key point information includes but is not limited to key points such as head, shoulders, elbows, wrists, hips, knees, and ankles.

[0042] Furthermore, the human pose estimation model uses a large-scale human pose dataset, such as COCO, MPII, or PoseTrack, during the training process to ensure the robustness of the model in different scenarios. In addition, the pose key point information can further include features such as the relative distance, angle, or movement trajectory between key points to enhance the accuracy and diversity of grouping.

[0043] Specifically, the method for grouping a series of pictures in S2 includes extracting the feature of the pose key point information of each picture, and performing clustering analysis according to the extracted features, and dividing pictures with similar poses into the same group.

[0044] Further, the clustering analysis algorithm includes but is not limited to K-means, hierarchical clustering, or DBSCAN, and the feature extraction method includes but is not limited to principal component analysis, linear discriminant analysis, or t-distributed stochastic neighbor embedding. In addition, the grouping method can be dynamically adjusted according to the complexity of the pose to ensure high similarity of the pictures within each group while avoiding overlap between groups.

[0045] Specifically, the mapping relationship in S3 is a one-to-one correspondence, and further includes encrypting the mapping relationship to prevent unauthorized access or tampering.

[0046] Further, the encryption process includes but is not limited to hash functions, digital signatures, or symmetric encryption algorithms. The mapping relationship can be stored in an encrypted mapping table and accessed through a key. In addition, the mapping relationship can be dynamically adjusted according to the length and complexity of the secret information to ensure the flexibility and security of the mapping.

[0047] Specifically, the diffusion model in S4 is a deep learning-based generative model, including but not limited to DDPM or DDIM, and the conditional input includes at least one of pose key point information, image style, and lighting conditions.

[0048] Further, the diffusion model uses a large-scale human image dataset such as DeepFashion, LSP, or Human3.6M during training to ensure the authenticity and diversity of the generated images. In addition, the conditional input can further include background information, texture features, or color distribution, etc., to enhance the quality and consistency of the generated images.

[0049] Specifically, the reduction method in S5 includes but is not limited to image cropping, image scaling, or image compression, and further includes encrypting or watermarking the reduced image to enhance the security of the image.

[0050] Further, the encryption process includes but is not limited to least significant bit steganography, discrete cosine transform steganography, or frequency domain steganography, and the watermarking process includes but is not limited to digital watermarking, text watermarking, or image watermarking. In addition, the reduction method can be adaptively adjusted according to the resolution and quality of the image to ensure that the reduced image maintains clarity while reducing the storage space and transmission bandwidth requirements.

[0051] Specifically, the inverse process in S6 includes extracting the pose key point information from the received image, grouping the image according to the extracted key point information, and restoring the corresponding binary sequence according to the grouping result, and finally converting the binary sequence into the original secret information.

[0052] Further, the method for extracting pose key point information includes, but is not limited to, key point detection algorithms based on deep learning, key point localization algorithms based on geometry, or key point tracking algorithms based on feature matching. In addition, the inverse process may further include decrypting or decompressing the extracted binary sequence to restore the original text, numerical, or symbolic information.

[0053] Specifically, the error detection and correction processing includes, but is not limited to, CRC (Cyclic Redundancy Check) or ECC (Error Correction Code) technology.

[0054] Further, the error detection and correction processing can be performed in real time during the transmission of the secret information to ensure the integrity and accuracy of the information. At the same time, the error detection and correction processing can be adaptively adjusted according to the noise level and bit error rate of the transmission channel to minimize the error rate during the information transmission process.

[0055] Embodiment 1: Text information hiding based on a diffusion model and human pose estimation. Assume that a confidential text message (such as "Project XYZ is scheduled for launch on December 15th") needs to be hidden in a set of human body images.

[0056] The specific implementation steps are as follows:

[0057] S1. Secret information processing: Convert the text message "Project XYZ is scheduled for launch on December 15th" into a binary sequence. For example, use ASCII encoding to convert each character into an 8-bit binary number, and segment the binary sequence with each segment having a length of 16 bits for subsequent mapping.

[0058] S2. Human pose estimation and grouping: Use the OpenPose model to extract pose key points from a set of human body images, extracting key point information such as the head, shoulders, elbows, wrists, hips, knees, and ankles of each image. Use the K-means clustering algorithm to group the images, classifying images with similar poses into the same group. For example, classify images of standing poses into one group and images of sitting poses into another group.

[0059] S3. Establishment of mapping relationship: Establish a one-to-one correspondence between the segmented binary sequence and the image groups. For example, the first binary sequence "0101010101010101" corresponds to the first group of images, and the second binary sequence "1010101010101010" corresponds to the second group of images.

[0060] S4. Diffusion model generates images: Use the DDPM model to generate images that conform to the pose keypoint information. For example, generate a new set of standing pose images based on the standing poses of the first group of images, and embed the corresponding binary sequence during the generation process;

[0061] S5. Image shrinking and stitching: Perform shrinking processing on the generated images. For example, reduce the resolution of each image from 1024x1024 to 256x256, and stitch multiple images into a large image;

[0062] S6. Information extraction: After the receiver receives the stitched image, use the OpenPose model to extract the pose keypoint information of each image, and restore the corresponding binary sequence according to the grouping result. Finally, convert the binary sequence into the original text information.

[0063] Example 2: Digital information hiding based on diffusion model and human pose estimation. Assume that a set of confidential digital information (such as "1234567890") needs to be hidden in a set of human body images.

[0064] The specific implementation steps are as follows:

[0065] S1. Secret information processing: Convert the digital information "1234567890" into a binary sequence. For example, use BCD encoding to convert each digit into a 4-bit binary number, and perform encryption processing on the binary sequence. Use the AES algorithm to encrypt the binary sequence;

[0066] S2. Human pose estimation and grouping: Use the AlphaPose model to extract pose keypoints from a set of human body images, extract the keypoint information of the head, shoulders, elbows, wrists, hips, knees, ankles, etc. of each image, and use the hierarchical clustering algorithm to group the images. Group the images with similar poses into the same group. For example, group the images with running poses into one group, and group the images with jumping poses into another group;

[0067] S3. Establish mapping relationship: Establish a one-to-one correspondence between the encrypted binary sequence and the image group. For example, the first binary sequence "0001001000110100" corresponds to the first group of images, and the second binary sequence "0101010101010101" corresponds to the second group of images;

[0068] S4. Diffusion model generates images: Use the DDIM model to generate images that conform to the pose keypoint information. For example, generate a new set of running pose images based on the running poses of the first group of images, and embed the corresponding binary sequence during the generation process;

[0069] S5. Image Shrinking and Stitching: Shrink the generated images, for example, reduce the resolution of each image from 1024x1024 to 512x512, and stitch multiple images into a large image;

[0070] S6. Information Extraction: After the receiver receives the stitched image, use the AlphaPose model to extract the pose key point information of each image, restore the corresponding binary sequence according to the grouping result, and finally decrypt the binary sequence into the original digital information.

[0071] Example 3: Symbol Information Hiding Based on Diffusion Model and Human Pose Estimation. Suppose a set of confidential symbol information (such as "@#$%^&*") needs to be hidden in a set of human images.

[0072] The specific implementation steps are as follows:

[0073] S1. Secret Information Processing: Convert the symbol information "@#$%^&*" into a binary sequence. For example, use Unicode encoding to convert each symbol into a 16-bit binary number, and perform compression processing on the binary sequence. Use the LZ77 algorithm to compress the binary sequence;

[0074] S2. Human Pose Estimation and Grouping: Use the HRNet model to extract the pose key point information of a set of human images, extract the key point information of the head, shoulders, elbows, wrists, hips, knees, ankles, etc. of each image, and use the DBSCAN clustering algorithm to group the images. Group the images with similar poses into the same group. For example, group the images of dancing poses into one group and the images of basketball-playing poses into another group;

[0075] S3. Establishing Mapping Relationship: Establish a one-to-one correspondence between the compressed binary sequence and the image group. For example, the first binary sequence "1101101011011010" corresponds to the first group of images, and the second binary sequence "1010110110101101" corresponds to the second group of images;

[0076] S4. Generating Images with Diffusion Model: Use the DDPM model to generate images that conform to the pose key point information. For example, generate a new set of dancing pose images according to the dancing poses of the first group of images, and embed the corresponding binary sequence during the generation process;

[0077] S5. Image Shrinking and Stitching: Shrink the generated images, for example, reduce the resolution of each image from 1024x1024 to 128x128, and stitch multiple images into a large image;

[0078] S6. Information extraction: After the receiver receives the spliced image, the HRNet model is used to extract the pose key-point information of each image, and the corresponding binary sequence is restored according to the grouping result. Finally, the binary sequence is decompressed into the original symbol information.

[0079] In summary, for the carrier-free image steganography based on the diffusion model and human pose estimation, through the technologies based on the diffusion model and human pose estimation, the steganography of the present invention can embed more secret information in a single picture. The diffusion model can generate high-quality and diverse images, while human pose estimation provides rich pose key-point information, which can be used as an encryption carrier. Compared with traditional steganography, the present invention can significantly improve the encrypted information capacity of a single picture without significantly increasing the image complexity.

[0080] Moreover, by grouping images through human pose estimation, multiple pictures are kept basically consistent in the theme content, and only differ in the actions of the characters. This design not only improves the visual consistency of the images, but also enhances the concealment of the steganography. Since multiple pictures are visually similar, it is difficult for attackers to discover the hidden secret information through simple visual analysis, thereby improving the confidentiality of the system. In addition, the invention can also enhance confidentiality and robustness, improve flexibility and scalability, and optimize information transmission and storage efficiency, realizing efficient, secure and concealed information hiding and transmission, and solving the problem that the prior art has low image generation quality due to technical limitations, specifically reflected in the insufficient peak signal-to-noise ratio, resulting in the current technical level of the diffusion model being unable to embed binary hash sequences or embed text like traditional models such as GAN (Generative Adversarial Network) to achieve the transmission of secret information, and can only transmit secret images with extremely small encryption capacity, making it difficult to achieve large-scale information transmission. At the same time, due to the label requirements of the generated images, the public key semantics are relatively similar, and potential attackers can guess the private key semantics from the public key, resulting in insufficient security.

[0081] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising a..." does not exclude the existence of additional identical elements in the process, method, article or device comprising the element.

[0082] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A carrier-free image steganography based on diffusion model and human posture estimation, characterized in that: The following steps are involved: S1, convert the secret information into a binary sequence according to the ASCII code, and wait to be encrypted as a secret sequence; S2. Extract a large number of full-body portraits from the Internet, extract the posture key point information of the target image based on the pre-trained human posture estimation model, group the series of pictures based on the key point arrangement features, such as the angle between the arm and the torso, and divide them into 2^n groups according to the binary bit number set for each picture; S3, mapping each group with a binary sequence, and then taking out the corresponding model from the corresponding group according to the binary sequence of the secret information; S4, inputting the posture key point information into the diffusion model as a condition to guide the model to generate an image consistent with the target posture; S5. Appropriately reduce the corresponding images to realize the splicing of multiple images, so that a large number of images in a long sequence can be compressed into an acceptable number of images; S6. The secret image is transmitted to the receiver, and the receiver extracts the secret information according to the inverse process.

2. According to claim 1, the carrier-free image steganography based on diffusion model and human posture estimation is characterized by: The secret information in S1 includes at least one of text, numbers or symbols, and further includes encrypting or compressing the secret information to improve the security of the information or reduce the storage space of the information.

3. The carrier-free image steganography based on diffusion model and human posture estimation according to claim 1 is characterized by: The human posture estimation model in S2 is a deep learning-based model, including but not limited to OpenPose, AlphaPose or HRNet, and the posture key point information includes but is not limited to key points such as head, shoulder, elbow, wrist, hip, knee, ankle, etc.

4. The carrier-free image steganography based on diffusion model and human posture estimation according to claim 1 is characterized by: The method for grouping the series of pictures in S2 includes extracting features from the posture key point information of each picture, and performing cluster analysis based on the extracted features to group pictures with similar postures into the same group.

5. The carrier-free image steganography based on diffusion model and human posture estimation according to claim 1 is characterized by: The mapping relationship in S3 is a one-to-one correspondence, and further includes encrypting the mapping relationship to prevent unauthorized access or tampering.

6. The carrier-free image steganography based on diffusion model and human posture estimation according to claim 1 is characterized by: The diffusion model in S4 is a generative model based on deep learning, including but not limited to DDPM or DDIM, and the conditional input includes at least one of posture key point information, image style, and lighting conditions.

7. The carrier-free image steganography based on diffusion model and human posture estimation according to claim 1 is characterized by: The reduction method in S5 includes but is not limited to image cropping, image scaling or image compression, and further includes encrypting or watermarking the reduced image to enhance the security of the image.

8. The carrier-free image steganography based on diffusion model and human posture estimation according to claim 1 is characterized by: The inverse process in S6 includes extracting posture key point information from the received image, grouping the image according to the extracted key point information, restoring the corresponding binary sequence according to the grouping result, and finally converting the binary sequence into the original secret information.

9. The carrier-free image steganography based on diffusion model and human posture estimation according to claim 2 is characterized by: The error detection and correction process includes but is not limited to CRC cyclic redundancy check or ECC error correction code technology.

Images