Capacity-extensible high-robustness carrier-free image steganography method
The dual-surface relationship between the carrier image library and the secret information sequence is constructed through the implicit denoising diffusion probability model DDIM and SIFT features. Combined with the VP tree structure, the problem of limited hidden capacity and insufficient robustness in the steganography method without carrier is solved. , realizes a carrier-free image steganography method with high robustness and perceived security.
Patent Information
- Application Number
- CN202411065517.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-11-19
AI Technical Summary
The existing carrier-free image steganography methods have problems such as limited hidden capacity, difficulty in balancing robustness and perceived security, and it is difficult to effectively resist image processing attacks.
The vector image library is generated by the implicit denoising diffusion probability model DDIM, and the bijective relationship between the vector image library and the secret information sequence is constructed through SIFT features and K-Means clustering algorithm, and the image retrieval efficiency and resolution are improved through the VP tree structure to Enhance robustness.
It realizes high-rootability, carrier-free image steganography with scalable capacity, improves the accuracy of extracting secret information and anti-image processing attack capabilities, and enhances the security and behavioral security of the steganography method.
Smart Images

Figure CN118984355A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of network security technology, and in particular to a highly robust carrier-free image steganography method with scalable capacity. Background Art
[0002] Steganography can embed secret information into the redundant part of the carrier and transmit it through an open channel, covering up the fact that "secret information exists", and has become one of the important ways to safely and reliably transmit secret information in a network environment. The modification-based steganography algorithm embeds secret information by modifying the content or characteristics of the carrier. Although these minor changes will not cause visual suspicion, specially designed steganalysis algorithms can detect the steganographic traces. In contrast, carrierless steganography generates or maps secret information into a secret object. Since the secret object that transmits secret information is completely indistinguishable from the normal object that does not transmit secret information, carrierless steganography has higher security than modification steganography and is difficult to be detected by existing steganalysis methods. Therefore, the carrierless steganography method can provide a safer means of covert communication and plays an indispensable role in the fields of privacy protection and data security.
[0003] The carrierless image steganography algorithm can be divided into the generation-based carrierless image steganography algorithm and the mapping-based carrierless image steganography algorithm. The generation-based carrierless image steganography algorithm uses image generation models, such as generative adversarial networks, flow models, diffusion models, etc., to directly generate the secret information into the stego-image. Liu et al. proposed the first GAN-based carrierless image steganography method in 2017, using ACGAN and the category label mapped by the secret information to generate the stego-image. Subsequently, a series of carrierless image steganography methods based on GAN variants emerged, such as WGAN, which generates another normal and independent image from the secret image, and StarGAN, which generates new face images using the mapping between secret information and facial attributes. However, their hiding capabilities are relatively limited, and the accuracy of secret information extraction is not high. Zhou et al. used a robust secret information encoding method and the Glow model to generate stego-images, which improved the embedding capacity of the steganography method and the accurate extraction of secret information. However, the quality of the generated images based on the Flow model is not high, which is easy to arouse the suspicion of the detector. Wei et al. first developed a diffusion model StegoDiffusion for generative steganography, and then combined DCT transform and inverse transform based on StegoDiffusion to generate high-quality stego images. Peng et al. proposed a generative image steganography method StegaDDPM based on a denoising diffusion probability model DDPM, which uses probability distribution and information sampling to hide secret information in the generated image. Although the generated image has high undetectability, deep neural networks are highly sensitive to small perturbations of the input image, making it difficult to correctly extract secret information from the stego image based on the model.
[0004] The mapping-based carrier-free image steganography method establishes a mapping relationship between the image hash sequence and the secret information, transmits the secret information by selecting the corresponding carrier image, and designs a robust image hash algorithm to ensure the lossless recovery of the secret information. Zhou et al. first introduced the concept of carrier-free and generated a hash sequence and mapped the secret information by comparing the average strength of adjacent blocks, but it cannot resist most image processing attacks. In order to improve the robustness of steganography methods, Zheng et al. used the SIFT algorithm to construct a robust image hash sequence to hide secret information [Reference 1: Zheng S, Wang L, Ling B, et al. Coverless information hiding based on robust image hashing [C] / / Intelligent Computing Methodologies: 13th International Conference, ICIC 2017, Liverpool, UK, August 7-10, 2017, Proceedings, Part III 13. Springer International Publishing, 2017: 536-547.]; Yuan et al. combined the BOF algorithm based on SIFT features to unify the feature dimensions of the image and then map the secret information; Zhang et al. proposed a robust algorithm based on discrete cosine transform (DCT) and latent Dirichlet allocation (LDA) topic classification [Reference 2: Zhang X, Peng F, Long M. Robust coverless image steganography based on DCT and LDA topic classification [J]. IEEE Transactions on Multimedia, 2018, 20(12): 3223-3238.]; Liu et al. proposed secret information mapping based on discrete wavelet transform (DWT) sequence [Reference 3: Liu Q, Xiang X, Qin J, et al. Coverless steganography based on image retrieval of DenseNet features and DWT sequence mapping [J]. Knowledge-Based Systems, 2020, 192: 105375.] and DenseNet features for image retrieval; Luo et al. introduced the object detection network Faster RCNN to improve the ability of steganographic images to resist geometric attacks; Liu et al. improved the robustness of the steganographic method by transmitting disguised images of steganographic images; Meng et al. directly generated the hash sequence of the image through an end-to-end CNN model, and introduced attention mechanism and adversarial training to improve the robustness of the model [Reference 4: Meng L, Jiang X, Zhang Z, et al. A robust coverless imagesteganography based on an end-to-end hash generation model [J]. IEEE Transactions on Circuits and Systems for Video Technology, 2022.]. Although these methods improve the robustness of covert communication, there are still problems such as low hidden capacity and incomplete mapping of secret information. In addition, Yuan, Liu et al. selected steganographic images of different subject types to send to improve the robustness of steganography, but this also reduced the behavioral security of covert communication. . Summary of the invention
[0005] Existing carrier-free image steganography methods still have shortcomings such as the hidden capacity is limited by the size of the public data set they rely on, and the difficulty in balancing robustness and perceived security. In order to solve the above problems, the present invention proposes a highly robust carrier-free image steganography method with scalable capacity. The implicit denoising diffusion probability model DDIM is used to generate a carrier image library in which the image types are consistent, which improves the hidden capacity and perceived security; at the same time, the image and the secret information sequence are made to correspond one to one through parallel mapping, and the designed bijective algorithm indirectly improves the hidden capacity; in order to improve the robustness of image processing attacks, the resolution of the image is enhanced before the stego image is sent.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions:
[0007] A highly robust carrier-free image steganography method with scalable capacity, comprising:
[0008] Step 1: Generate N face images using the implicit denoising diffusion probability model DDIM and standardize the image size to obtain the face image library Face_GID;
[0009] Step 2: Determine the length l of the secret information sequence, And divide the secret information of length L to be sent into num secret information sequences, in, represents the round up function;
[0010] Step 3: Extract the SIFT features of the face image, calculate the centroid features using the K-Means clustering algorithm, sort the images of Face_GID in ascending order according to the Euclidean distance from the SIFT features to the centroid, and obtain the index number of each image in Face_GID;
[0011] Step 4: Determine the carrier image based on the length l of the secret information sequence and the index numbers of each image of Face_GID in ascending order, and finally construct the carrier image library CF_GID;
[0012] Step 5: Parallel mapping of each carrier image in CF_GID with a l-bit binary secret information sequence to generate a mapping table, and selecting the corresponding carrier image as the stego-image according to the secret information sequence through the mapping table;
[0013] Step 6: Based on the resolution enhancement factor t, the resolution of the steganographic image is sequentially enhanced (eg, by an interpolation algorithm) and sent.
[0014] Furthermore, the steps 1 to 6 are performed at the sending end, and after the receiving end receives the steganographic image with improved resolution sent by the sending end, the following process is also included:
[0015] Step 7: Obtain the mapping table and carrier image library CF_GID through the implicit denoising diffusion probability model DDIM, the length of the secret information sequence l, and the shared parameters of the carrier image selection method, and obtain the retrieval image library RF_GID in combination with the resolution improvement factor t;
[0016] Step 8: Use the SIFT features of each image in RF_GID to construct the VP tree structure of RF_GID;
[0017] Step 9: Extract the SIFT features of the received image, retrieve the VP tree structure of RF_GID, and obtain the original stego image;
[0018] Step 10: Look up the mapping table based on the stego-image to obtain the mapped binary secret information sequence;
[0019] Step 11: Connect all secret information sequences in sequence to obtain the sent secret information.
[0020] Furthermore, in step 2, if L is not divisible by l, the tail of the last secret information sequence is padded with zeros to form a complete sequence, and the number of padded zeros is recorded in the last transmitted image.
[0021] Furthermore, in step 8, the VP tree structure of RF_GID is constructed in the following manner:
[0022] Step 8.1: The image farthest from the cluster centroid in RF_GID is recorded as point V, and the rest of the images are points V′;
[0023] Step 8.2: Calculate the SIFT feature distance d of all V' points from point V i ,i={1,2,…,m,…,n-1}, take the median d m As the radius to divide the non-dominant point, generate the left subtree V0 and right subtree V1 of the dominant point V. If d i <d m , then V′∈V0; otherwise, V′∈V1. After the division is completed, the non-dominant points farthest from the dominant point in the left subtree and the right subtree are taken as the dominant points of the two subtrees respectively;
[0024] Step 8.3: Repeat the process of step 8.2 until the subtree is empty. At this point, the VP tree structure of RF_GID is constructed.
[0025] Furthermore, in step 9, the VP tree structure of RF_GID is retrieved in the following manner:
[0026] Step 9.1: Set the distance threshold r to extract the SIFT features of the received image Search the VP tree of RF_GID from top to bottom;
[0027] Step 9.2: Calculate the distance from the received image to the vantage point. If Then add the advantage point to the neighbor point queue and use the distance to update the threshold. Then recursively search the left subtree, if Then recursively search the right subtree, if both and Two conditions, recursively search the left subtree and the right subtree;
[0028] Step 9.3: Select the nearest neighbor point in the queue as the stego-image corresponding to the received image.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] (1) A carrier-free image steganography scheme is proposed based on an image generation model to construct a carrier library. Generating a carrier image library can better guarantee the capacity requirements and behavioral security of such steganography methods.
[0031] (2) A bijection algorithm between the carrier image library and the secret information sequence is designed. This achieves full coverage of the carrier image library on the secret information sequence, indirectly improving the capacity of the steganography method.
[0032] (3) The advantages of SIFT features in identical image matching and retrieval are verified. Experimental results show that compared with the deep features of the neural network DenseNet, the SIFT features of the preprocessed image are more conducive to the matching and retrieval of identical images, and strongly guarantee the accuracy of secret information extraction. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 A flowchart of a highly robust carrier-free image steganography method with scalable capacity according to an embodiment of the present invention;
[0034] Figure 2 The same image matching success rate based on SIFT and DenseNet121 provided in the embodiment of the present invention;
[0035] Figure 3 The matching success rate of the same image at different resolutions based on SIFT provided in the embodiment of the present invention. DETAILED DESCRIPTION
[0036] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:
[0037] like Figure 1 As shown in the figure, a highly robust and scalable carrier-free image steganography method is proposed. The process of hiding secret information at the sender is divided into six steps, as follows:
[0038] S101: Generate N face images I={I i |i=1,2,…,N}, and the size is standardized to M×M, and the image library is denoted as Face_GID.
[0039] S102: Determine the length of the secret information sequence The secret information of length L to be sent is divided into a secret information sequence SI = {s j |j=1,2,…,num}, in, If L is not divisible by l, the last secret information sequence is padded with zeros to form a complete sequence, and the last transmitted image is used to record the number of padded zeros.
[0040] S103: Extracting image I i SIFT features are used to calculate the centroid feature F using the K-Means clustering algorithm. c , Arrange the images of Face_GID in ascending order according to the Euclidean distance from the SIFT feature to the centroid, and obtain the index number ID of each image in Face_GID.
[0041] Fc =Kmeans(F1,F2,…,F N )#(1)
[0042] ID i =ID(I i )=argsort(d1,d2,…,d N )#(2)
[0043] Among them, d i =ED(F i ,F c ), Kmeans(), argsort(), and ED() are clustering algorithm function, ascending sort function, and Euclidean distance function respectively.
[0044] S104: Determine the carrier image C based on the length l of the secret information sequence and the index numbers of the images of Face_GID arranged in ascending order k , build the carrier image library CF_GID.
[0045] CF_GID = {C k |k=1,2,…,2 l}#(3)
[0046] C k =I i #(4)
[0047] Where k = ID (I i ).
[0048] S105: Parallel mapping of each carrier image in CF_GID with a l-bit binary secret information sequence to generate a mapping table. j , select the corresponding carrier image C k As the stego-image S j .
[0049] S j =C k #(5)
[0050] Where k = binary2decimal(s j ), binary2decimal() converts a binary sequence into a decimal number.
[0051] S106: Based on the resolution enhancement factor t, the resolution of the steganographic image is sequentially enhanced and sent.
[0052] The process of extracting secret information at the receiving end is divided into five steps, as follows:
[0053] S107: The mapping table and carrier image library CF_GID are obtained through the shared parameters of the image generation model DDIM, the secret information sequence length l, the carrier image selection method, etc., and the retrieval image library RF_GID is obtained by combining the resolution improvement factor t.
[0054] S108: Using the SIFT features of each image in the search image library RF_GID, a VP tree structure of RF_GID is constructed.
[0055] S109: Extract SIFT features of the received image, and obtain the original steganographic image with the help of the VP tree structure of RF_GID.
[0056] S110: Searching a mapping table based on the stego-image to obtain a mapped binary secret information sequence.
[0057] S111: Connect all secret information sequences in sequence to obtain the sent secret information.
[0058] It is worth noting that the received image referred to in the present invention is a steganographic image subjected to an image processing attack, so as to be distinguished from the original steganographic image.
[0059] The method specifically includes:
[0060] A. Generate an image library
[0061] The hiding capacity of the image steganography without carrier based on mapping depends on the number of images in the image library. The number of images in the currently public image datasets is fixed, which limits the peak value of the hiding capacity. In addition, the random selection of carrier images will seriously affect the security of the image steganography without carrier. The inconsistent subject types of the carrier images make the visual perception of the image obviously abnormal, and the behavioral security of the image transmission cannot be ensured. The present invention proposes to construct a carrier image library using an image generation model, which not only realizes the controllable number of carrier images, but also meets the requirement that all carrier images belong to the same category.
[0062] First, select an image generation model. The image generation model used in the proposed method is the current mainstream implicit denoising diffusion probability model DDIM, which generates images of comparable quality to the generative adversarial network model GAN and does not require adversarial training. In addition, compared with the original denoising diffusion probability model DDPM, DDIM generates images of better quality and faster speed.
[0063] Secondly, determine the number of carrier images. Using the image generation model DDIM, any number of images can be generated. However, the more images there are, the longer the secret information sequence that can be mapped to each image, but it will also lead to low efficiency in image retrieval. Therefore, the number of images generated needs to balance the length of the mapped secret information and the retrieval efficiency of the image. The present invention generates 10,000 images for experimental testing.
[0064] Finally, the carrier image is generated. Based on the training dataset CelebA and 10,000 random seeds, the proposed method generates 10,000 64×64 face images with a step size of step=20, and finally constructs a carrier image library.
[0065] B. Designing a bijective algorithm
[0066] A good mapping algorithm can achieve a one-to-one correspondence between the carrier image and the secret information sequence, and realize the bijection of the secret information sequence. The present invention uses the image SIFT feature and the K-Means clustering algorithm to design a bijection algorithm between the carrier image and the secret information sequence, such as Figure 1 as shown in .
[0067] First, the SIFT features of each image in the generated image library are extracted, and the centroid features of the image library are obtained using the K-Means clustering algorithm. Then, the Euclidean distance from each image to the centroid is calculated based on the features, and the images are arranged in ascending order of distance. Finally, the sorted images and secret information sequences are mapped in parallel, and a mapping table is generated. The mapping table contains three columns of data, namely the name of the image, the mapped secret information sequence, and the storage address of the image. During the clustering process, the K-Means algorithm uses random seeds to determine the initial centroid image.
[0068] The mapping algorithm of the present invention does not directly use image features to map secret information sequences, but reorders images with the help of image features. The advantage of this design is that there is no problem of incomplete mapping, such as the secret information sequence cannot find the mapping image, and multiple images repeatedly map the same secret information sequence.
[0069] C. Constructing VP tree
[0070] In the carrier-free image steganography scheme proposed in this invention, the receiving end needs to retrieve the original stego image from the image library based on the received image. Considering that it takes a lot of time to match images one by one, in order to improve the retrieval efficiency, the proposed method constructs the image library VP tree structure based on the similarity between the image SIFT features.
[0071] VP tree is a high-dimensional index structure based on distance metric space. It can effectively process large-scale data sets and takes up less memory. It uses the idea of binary search to divide the feature space by the distance information between the advantage point V and the non-advantage point V′. Ideally, the retrieval complexity of a VP tree with n feature points is O(logn).
[0072] The steps of constructing the VP tree of the image library in the proposed method are as follows: First, the image farthest from the cluster centroid in the image library RF_GID is recorded as point V, and the rest of the images are all points V′. Second, the SIFT feature distance d of all points V′ from the points V is calculated. i ,i={1,2,…,m,…,n-1}, take the median d m As the radius to divide the non-dominant point, generate the left subtree V0 and right subtree V1 of the dominant point V. i <d m , then V′∈V0; otherwise, V′∈V1. After the division is completed, the non-dominant points farthest from the dominant point in the left subtree and the right subtree are used as the dominant points of the two subtrees. The third step is to repeat the process of the second step until the subtree is empty. At this point, the generated VP tree covers the entire image library, and the VP tree construction is completed.
[0073] The steps to retrieve the VP tree are as follows: Step 1, set the distance threshold r and extract the SIFT features of the received image Search the VP tree from top to bottom. The second step is to calculate the distance from the received image to the vantage point. If Then add the advantage point to the neighboring point queue and use the distance to update the threshold r. Then recursively search the left subtree. Then recursively search the right subtree. If both and If the two conditions are met, the left subtree and the right subtree are recursively retrieved. In the third step, the nearest neighbor in the queue is selected as the stego-image corresponding to the received image.
[0074] D. SIFT feature extraction and resolution improvement verification
[0075] The robustness of image steganography without a carrier depends on the similarity between the received stego-image and the original stego-image. The higher the similarity, the greater the accuracy of extracting the secret information. Therefore, how to retrieve the original stego-image in the image library based on the received stego-image is the key to improving robustness. In addition, the higher the resolution of the image, the stronger its ability to resist image processing attacks.
[0076] In order to verify the influence of feature extraction algorithm and image resolution on image matching, the present invention designed two experiments based on the test dataset CelebA*. The first experiment tested the image matching success rate based on SIFT method and deep model DenseNet121 method at the same resolution. The experimental results are shown in Figure 1 ; The second experiment tested the image matching success rate based on SIFT method at different resolutions. The experimental results are shown in Figure 2The datasets and numbers (01-27) in the figure can be found in the parameter setting description in the experimental section. In addition, the features used in image matching are all unprocessed SIFT and DenseNet121 original features.
[0077] from Figure 2 It can be seen that the matching success rate of the same image based on SIFT features is closer to the outer circle of 100%, and the result curve is smoother, indicating that SIFT features are more stable and robust in various image processing attacks. Therefore, we use traditional SIFT features for image retrieval in the proposed method.
[0078] Figure 3 It is intuitively shown that when the resolution is increased to twice the original resolution, the success rate of the same image matching based on SIFT features is enhanced. Therefore, the present invention performs resolution enhancement processing on the sent steganographic image to enhance the robustness of the steganographic method.
[0079] E. Experimental Results and Analysis
[0080] In order to verify the effectiveness of the proposed method, the present invention conducted experimental tests on hiding capacity and robustness. Based on the consideration of the same category of carrier images and SIFT features in the proposed method, the present invention selected and reproduced the four most advanced methods of this type of steganography methods [Reference 1], [Reference 2], [Reference 3], [Reference 4] for experimental comparison and result analysis.
[0081] E.1 Experimental Setup
[0082] The test data set of the present invention includes ImageNet (500 128×128 images), CelebA (500 64×64 images) and Face_GID (500 64×64 images), which are marked as ImageNet*, CelebA* and Face_GID*, respectively. At the same time, the hash sequence length is set to 8. According to the hash sequence generation rules in each comparison method, the image in method [Document 1] is divided into 4 blocks and an 8-bit hash sequence is generated. The image in methods [Document 2] and [Document 3] is divided into 16 sub-blocks and the first 8 bits are taken as the final hash sequence of the image. The above hash sequence generation processes all adopt a row-first-column-later method to scan the image sub-blocks. In addition, Documents 2 and 3 both use full-size images for feature extraction. When reproducing method [Document 4], the present invention has epoch=200, learning rate lr=0.00001, and the loss function remains consistent with the original text.
[0083] E.2 Hidden Capacity
[0084] Different from modification-based image steganography, the stegocapacity of mapping-based carrierless image steganography is the number of images N required to hide a fixed-length secret message. s It is related to the total length L of the secret information and the length l of the secret information sequence mapped by a single image, and is expressed as follows:
[0085]
[0086] From formula (6), we can see that the number of stego images N s As the length of the secret information sequence l increases, it decreases. According to the experimental parameter setting l = 8, when L = 100B, N s = 101. In order to avoid the situation where L cannot be divided by l, an image is added to record the number of 0s appended to the last secret information sequence.
[0087] However, formula (6) only gives the number of images that need to be sent when hiding secret information of a fixed length from a theoretical calculation. Its premise is that any secret information sequence has a corresponding carrier image. But in the actual information hiding process, the designed mapping algorithm cannot guarantee the establishment of this premise. Based on four test data sets and l=8, the present invention tests the number of different hash sequences that can be generated by the mapping algorithms of the comparison method and the proposed method to verify the possibility of the premise being established.
[0088] Table 1 Number of unique values of hash sequences generated by different methods
[0089]
[0090] As can be seen from Table 1, only the proposed method can generate a unique hash sequence for each image, satisfying the above premise. For methods [Reference 1], [Reference 2], [Reference 3] and [Reference 4], since there are multiple images generating the same hash sequence, it is necessary to first select images to build a carrier image library and then perform information hiding, so that the number of stego images sent can meet the calculation result of formula (6).
[0091] E.3 Robustness
[0092] During the transmission process, the stego image will be subjected to a variety of image processing attacks such as JPEG compression and noise, which inevitably destroy the content of the image. The robustness of the steganographic method studies the key issue of how to resist these image processing attacks to improve the accuracy of secret information extraction. The proposed method takes into account common image processing attacks, and the specific parameters are shown in Table 2.
[0093] Table 2 Common image processing attacks and parameter settings
[0094]
[0095] In the steganography method, the robustness evaluation index is the accuracy of extracting the secret information. Assume that the secret information sequence hidden in the i-th sent image is ts i , the secret information sequence extracted from the i-th received image is rs i , the number of stego images is N s , then the formula for extracting the correct rate AC is defined as:
[0096]
[0097] Among the 27 image processing attacks set by the present invention, the robustness test results of the four comparison methods and the proposed method are shown in Tables 3 to 5.
[0098] Table 3 Robustness comparison of different methods on the ImageNet* dataset
[0099]
[0100] Table 4 Robustness comparison of different methods on the dataset CelebA*
[0101]
[0102]
[0103] Table 5 Robustness comparison of different methods on the Face_GID* dataset
[0104]
[0105] The experimental results in Tables 3-5 show that the robustness of the proposed method is better than that of the four comparison methods in almost all image processing attacks. There are two reasons for this. One is due to the way of extracting secret information. The proposed method does not directly use the received image to extract secret information, but retrieves the original steganographic image through the received image, and then uses the mapping table to query the corresponding secret information. The other is due to the hash algorithm. The proposed method maps the secret information according to the sorting position of the steganographic image in the carrier image library, rather than using image features for hash mapping.
[0106] Among the four comparison methods, method [1] has weak resistance to various image processing attacks, and the experimental results are lower than those of methods [2] and [3] that also use traditional image feature extraction algorithms and image segmentation. This is because method [1] has the least number of image segments, and each image sub-block generates a 2-bit hash sequence, while methods [2] and [3] each have a 1-bit hash sequence. Image processing attacks have a greater impact on the hash sequence of the image sub-block in method [1]. In other words, for hash sequence generation algorithms based on image segmentation, the more sub-blocks the image is divided into, the more robust the hash algorithm is. In addition, methods [2] and [3] are relatively weak in resistance to geometric attacks such as cropping, rotation, and translation. Method [4] has higher resistance to rotation and translation attacks than the other three comparison methods, which is related to the attention mechanism in the method. Moreover, as the resolution of the carrier image increases, the average robustness of method [4] becomes stronger.
[0107] E.4 Security
[0108] The security of the carrierless image steganography based on mapping can be analyzed from two aspects, namely feature detection and visual perception. In terms of image feature detection, the transmission of secret information is realized by selecting the carrier image, and the content and structure of the image are not modified. Therefore, the detector cannot identify the stego image by statistically analyzing the features of the image, which fundamentally has the ability to resist steganalysis. From this aspect, carrierless image steganography is safe. In terms of visual perception, the carrier image library used by the proposed method is of the same category, and there is no situation where the categories of the stego images are greatly different, so it will not easily arouse the suspicion of the detector, so from the perspective of perception, the proposed method is safe.
[0109] In summary, the present invention proposes a highly robust carrier-free image steganography method based on a generated image library. The proposed method uses the image generation model DDIM to construct a carrier image library. The flexible and controllable number of images allows the hidden capacity to be arbitrarily expanded, and the generated images of the same category have high perceived security. The images are sorted according to the image SIFT features and the K-Means clustering algorithm, and the image positions are parallelly mapped to the secret information sequence to achieve a bijective state. The VP tree structure of the carrier image library is constructed. Based on this structure, the original stego image is retrieved by receiving the image, achieving a high retrieval success rate and greatly enhancing the robustness of the steganography method. Experimental results show that compared with existing steganography methods of this type, the proposed method has stronger robustness and a more easily expandable hidden capacity.
[0110] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.
Claims
1. A highly robust carrier-free image steganography method with scalable capacity, characterized in that: include: Step 1: Generate N face images using the implicit denoising diffusion probability model DDIM and standardize the image size to obtain the face image library Face_GID; Step 2: Determine the length l of the secret information sequence, And divide the secret information of length L to be sent into num secret information sequences, in, represents the round up function; Step 3: Extract the SIFT features of the face image, calculate the centroid features using the K-Means clustering algorithm, sort the images of Face_GID in ascending order according to the Euclidean distance from the SIFT features to the centroid, and obtain the index number of each image in Face_GID; Step 4: Determine the carrier image based on the length l of the secret information sequence and the index numbers of each image of Face_GID in ascending order, and finally construct the carrier image library CF_GID; Step 5: Parallel mapping of each carrier image in CF_GID with a l-bit binary secret information sequence to generate a mapping table, and selecting the corresponding carrier image as the stego-image according to the secret information sequence through the mapping table; Step 6: Based on the resolution enhancement factor t, the resolution of the steganographic image is sequentially enhanced and sent.
2. A highly robust carrier-free image steganography method with scalable capacity according to claim 1, characterized in that: The steps 1 to 6 are performed at the sending end. After the receiving end receives the steganographic image with improved resolution sent by the sending end, the following process is also included: Step 7: Obtain the mapping table and carrier image library CF_GID through the implicit denoising diffusion probability model DDIM, the length of the secret information sequence l, and the shared parameters of the carrier image selection method, and obtain the retrieval image library RF_GID in combination with the resolution improvement factor t; Step 8: Use the SIFT features of each image in RF_GID to construct the VP tree structure of RF_GID; Step 9: Extract the SIFT features of the received image, retrieve the VP tree structure of RF_GID, and obtain the original stego image; Step 10: Look up the mapping table based on the stego-image to obtain the mapped binary secret information sequence; Step 11: Connect all secret information sequences in sequence to obtain the sent secret information.
3. A highly robust carrier-free image steganography method with scalable capacity according to claim 1, characterized in that: In step 2, if L is not divisible by l, the tail of the last secret information sequence is padded with zeros to form a complete sequence, and the number of padded zeros is recorded in the last transmitted image.
4. A highly robust carrier-free image steganography method with scalable capacity according to claim 2, characterized in that: In step 8, the VP tree structure of RF_GID is constructed in the following manner: Step 8.1: The image farthest from the cluster centroid in RF_GID is recorded as point V, and the rest of the images are points V′; Step 8.2: Calculate the SIFT feature distance d of all V' points from point V i ,i={1,2,…,m,…,n-1}, take the median d m As the radius to divide the non-dominant point, generate the left subtree V0 and right subtree V1 of the dominant point V. If d i <d m , then V′∈V0; otherwise, V′∈V1. After the division is completed, the non-dominant points farthest from the dominant point in the left subtree and the right subtree are taken as the dominant points of the two subtrees respectively; Step 8.3: Repeat the process of step 8.2 until the subtree is empty. At this point, the VP tree structure of RF_GID is constructed.
5. A highly robust carrier-free image steganography method with scalable capacity according to claim 4, characterized in that: In step 9, the VP tree structure of RF_GID is retrieved in the following manner: Step 9.1: Set the distance threshold r to extract the SIFT features of the received image Search the VP tree of RF_GID from top to bottom; Step 9.2: Calculate the distance from the received image to the vantage point. If Then add the advantage point to the neighbor point queue and use the distance to update the threshold. Then recursively search the left subtree, if Then recursively search the right subtree, if both and Two conditions, recursively search the left subtree and the right subtree; Step 9.3: Select the nearest neighbor point in the queue as the stego-image corresponding to the received image.
