Image-video cross-dimensional steganography method, system and medium

Abstract

This invention provides a method, system, and medium for cross-dimensional image-video steganography. The method comprises: segmenting a secret image into K secret image blocks; inputting the preprocessed secret image blocks and carrier video frames into a MemoryBank module, selecting K optimal carrier video frames for information hiding based on feature matching, and ensuring that no duplicates are selected between frames; inputting the K secret image blocks and the selected carrier video frames into a VSHI-Net model for steganalysis processing, generating an output video frame containing the secret image blocks; and using the video frame containing the secret image block as one of the inputs to the VSHI-Net model. After network processing, the recovered secret image blocks are generated, and the recovered secret image is obtained through image splicing and reassembly. This invention achieves efficient cross-dimensional embedding and lossless recovery of the secret image into the carrier video, improving the carrier utilization, concealment, and robustness of cross-modal steganography.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision, deep learning and information steganography, and in particular to an image-video cross-dimensional steganography method, system and medium based on a reversible neural network. Background Art

[0002] Against the backdrop of today's rapidly developing information technology, steganography, as a covert communication method, has gradually attracted widespread attention. The basic principle of steganography is to embed secret information in other data carriers (such as images, audio, and video) in an unobtrusive manner, ensuring that the content is not detected by third parties during transmission. Unlike traditional encryption techniques, steganography aims not only to protect the content of information from prying eyes but also to conceal the existence of the information itself, thereby ensuring the privacy and security of both communicating parties. With the popularization of digital media and the advancement of computing power, steganography has shown significant application potential in a variety of fields, including network security, copyright protection, and identity authentication.

[0003] Traditional video steganography techniques primarily cover three domains: spatial, frequency, and compression domains. In the spatial domain, the least significant bit (LSB) technique is widely used. This technique replaces the n least significant bits of a carrier image pixel with the n most significant bits of the secret data. Frequency domain methods utilize techniques such as the discrete cosine transform (DCT) and wavelet transform (DWT) to embed secret data within transformed video coefficients. Compression domain techniques rely on video coding standards such as H.264 / AVC and H.265 / HEVC for steganography. However, these traditional methods, due to their limited hiding capacity and insufficient concealment, no longer meet current application requirements.

[0004] To overcome the limitations of traditional steganography techniques, deep learning has been gradually applied to this field. Early research proposed an end-to-end image steganography framework based on deep learning, enabling the bidirectional embedding and extraction of full-size secret images in a carrier. Subsequently, this technology was extended to the video field, forming a spatiotemporal data hiding paradigm using continuous video frames as carriers. During this process, generative adversarial networks (GANs) became a mainstream solution due to their ability to improve steganography security. Their typical architecture enhances the visual imperceptibility of the carrier by collaboratively optimizing the generator and discriminator modules. Subsequent research further strengthened the model's robustness by improving the network structure or designing new loss functions. With the development of reversible neural networks (INNs), their precisely reversible mathematical properties have provided a new path for information steganography. This technology requires only a single model to simultaneously achieve data hiding and lossless recovery. Currently, the technology has been technically verified in both multi-level steganography of static images and dynamic video steganography scenarios.

[0005] In the field of information steganography, although deep learning has significantly improved the ability to hide information in homogeneous carriers (such as image-image and video-video), when faced with cross-dimensional tasks such as embedding secret images into video carriers, due to the spatial structure of image data and the spatiotemporal dynamic characteristics of video carriers, existing methods still face fundamental technical bottlenecks in coordinating the expression of two-dimensional static features and three-dimensional motion features. These bottlenecks mainly focus on the following two aspects:

[0006] 1. System adaptability barriers caused by homogeneous carrier constraints

[0007] Most current mainstream steganography systems rely on a strict dimensionality match between the secret information and the carrier medium. In typical application scenarios, when transmitting secret information such as images, existing solutions (such as deep learning-based frameworks like SteganoGAN and HiDDeN) require the use of homogeneous image carriers for embedding. This technical paradigm exhibits significant limitations in multimodal collaboration scenarios. In real-world deployments, heterogeneous carrier resources are often encountered (for example, video surveillance systems primarily output video streams, while cloud storage platforms primarily store video files). This forces the system to perform dimensionality reduction preprocessing—breaking the video carrier into independent image frame sequences before performing traditional image steganography frame by frame. This process not only introduces additional computational burden but also fragments critical temporal information. When the carrier video undergoes compression or resolution conversion, the preprocessed frame sequence may experience pixel distortion and synchronization misalignment, further weakening the reliability of information extraction. This constraint also rigidifies the system architecture, preventing the dynamic selection of the optimal carrier type based on the environment, significantly limiting cross-platform deployment flexibility.

[0008] 2. Insufficient research on cross-domain steganography technology and its potential has not been fully realized

[0009] Although a small number of studies have begun to explore cross-domain hiding between images and videos (for example, hiding images in videos, or hiding videos in images), these methods are still in their early stages and have significant potential for improvement. Specifically, they are manifested in two aspects:

[0010] First, some methods still use traditional steganography algorithms with limited computational features, weak robustness and security. Their performance (such as core indicators such as carrier capacity, imperceptibility, and security) has shown a gap compared to the current mainstream deep learning-driven methods.

[0011] Second, while other methods utilize video as a carrier, their core logic simply involves extracting a single frame from the video to embed a secret image. This approach is essentially an extension or variation of the "image-in-image" model, treating video frames as independent sets of images and completely failing to deeply explore and utilize the inherent temporal dimension and high-dimensional properties of video. This wastes valuable carrier resources and fails to effectively leverage the additional redundancy provided by dynamic video to improve hiding capacity, robustness, or security. Summary of the Invention

[0012] The main purpose of the present invention is to provide an image-video cross-dimensional steganography method, system and medium based on a reversible neural network, aiming to fully explore and utilize the high-dimensional redundant information of the video carrier relative to the image, and realize cross-dimensional information steganography technology in which the carrier video hides the secret image.

[0013] In order to achieve the above objectives, the present invention proposes an image-video cross-dimensional steganography method, which comprises the following steps:

[0014] Step S10: In the forward hiding stage, the secret image is pre-processed by the sliding window operation of the image cropping module to divide the secret image into secret image blocks;

[0015] Step S20: The pre-processed secret image block and the carrier video frame are input into the MemoryBank module, and the pre-processed secret image block and the carrier video frame are filtered based on the feature matching degree. The optimal carrier video frames are used for information hiding and ensure that there is no repeated selection between frames;

[0016] Step S30: secret image blocks and the filtered The carrier video frames are input in parallel to the image-video cross-dimensional steganography model VSHI-Net based on the reversible hidden neural network for steganography processing, and the video frame output containing the secret image block is generated;

[0017] Step S40: In the reverse recovery phase, the video frame containing the secret image block is used as one of the inputs of the image-video cross-dimensional steganography model VSHI-Net based on a reversible hidden neural network. After network processing, the recovered secret image block is generated, and finally the recovered secret image is obtained by image splicing and reassembly.

[0018] A further technical solution of the present invention is that, in step S10, a sliding window is used to divide the secret image into blocks during the image cutting process. Since the size of the secret image block will inevitably be reduced after cropping, it will cause a size mismatch problem. Therefore, each cropped secret image block needs to be padded to restore the original size; in step S40, during the image stitching stage, the restored secret image block needs to first remove the padded area, and then be spliced ​​and reassembled according to the original spatial coordinates recorded during the cropping process, and finally reconstruct the complete secret image.

[0019] A further technical solution of the present invention is that step S20 includes:

[0020] Step S201: In the initialization phase, the secret image block sequence and the carrier video are input into the pre-trained ResNet network for feature extraction to obtain the secret image block feature set. and the carrier video frame feature set ;

[0021] Step S202, in the selection stage, uses the cosine similarity measurement method to perform sequential matching calculations on the features of each secret image block and all carrier video frame features, and selects the video frame with the highest matching degree as the steganographic carrier of the image block based on the similarity calculation results.

[0022] A further technical solution of the present invention is that the reversible hidden neural network in step S30 includes three parts: a discrete wavelet transform module, a reversible hidden module and an inverse wavelet transform module. In the forward hiding process, the input of the reversible hidden neural network includes three parts: a cropped secret image block sequence , optimally matched carrier video frame and Gaussian noise ;

[0023] The step S30 includes:

[0024] Step S301: The clipped secret image block sequence is transformed by a discrete wavelet transform module. , optimally matched carrier video frame Perform high and low frequency wavelet decomposition to generate secret image wavelet subbands respectively and carrier video wavelet subbands ;

[0025] Step S302: Generate secret image wavelet subband , carrier video wavelet subband With Gaussian noise Input the reversible hidden module for processing and output the wavelet subband containing the secret image block , wavelet domain information residual and noise residual ;

[0026] Step S303: wavelet subband , wavelet domain information residual After reconstruction by the inverse wavelet transform module, the video frame containing the secret image block is output and residual signal ;

[0027] Among them, the forward hidden process can be formally expressed as:

[0028] ;

[0029] The step S40 includes: introducing random sampling variables in the reverse recovery process , the carrier video frame containing the secret image block is transformed into With random sampling variables Decomposition into wavelet subbands and , while multiplexing Gaussian noise ;

[0030] Through the inversion operation of the reversible hidden module, the carrier video wavelet subband is restored , secret image wavelet subband and noise residual ;

[0031] The carrier video wavelet subband is transformed into , secret image wavelet subband Reconstruct into the spatial domain and obtain the restored carrier video frames With secret image block ;

[0032] The reverse recovery process can be formally expressed as:

[0033] ;

[0034] Forward mapping function and the inverse mapping function A set of shared network parameters.

[0035] A further technical solution of the present invention is that the reversible hidden module is composed of The first reversible hidden block input consists of three parts: the carrier video wavelet subband , secret image wavelet subband and Gaussian noise ,in, and Derived from the carrier video frame Frequency domain representation of the secret image block after processing by discrete wavelet transform module ;

[0036] ;

[0037] ;

[0038] Where, Represents the discrete wavelet transform operator; the output of the first reversible hidden block consists of three parts: the primary wavelet subband containing the secret image block , wavelet domain information residual and noise residual , its mathematical expression is as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] Where, Defined as a sigmoid activation function scaled by a constant factor, (·), (·), (·), (·), (·), (·), (·)and (·) is any nonlinear function that can be learned. reversible hidden blocks ( ), whose input is the wavelet subband containing the secret image block output in the previous stage , information residual and noise residual , after transformation in this stage, the output is updated to:

[0043] ;

[0044] ;

[0045] ;

[0046] Assume that the reversible hidden module contains reversible hidden blocks, after the After processing the reversible hidden blocks, the system will output the final wavelet subband containing the secret image block , cumulative information residual and noise residual ; These three output quantities correspond to the final state variables of the reversible hidden process:

[0047] ;

[0048] and are all subband signals in the wavelet domain, and The encrypted video frames and residual signals converted into the spatial domain need to be reconstructed in the frequency-space domain through the inverse wavelet transform module. Its mathematical expression is:

[0049] ;

[0050] ;

[0051] Where, Represents the inverse wavelet transform operator.

[0052] A further technical solution of the present invention is that the information flow direction of the reverse recovery process is opposite to that of the forward process, which is manifested as The reversible hidden block Reversible hidden block is transmitted backward; in the reverse recovery process, the system only has the video frame containing the secret image block As input, but lose information and noise residual Unavailable; for this reason, an auxiliary variable is introduced Alternative , and multiplexed Gaussian noise Alternative , in order to achieve the reversibility of hiding and restoring process, Must obey the same distribution; assuming that both obey Gaussian distribution, that is, ;

[0053] The specific implementation process of the reverse recovery process is as follows: Video frame containing secret image block and randomly sampled variables First, convert it to the wavelet domain through the discrete wavelet transform module to generate:

[0054] ;

[0055] Afterwards, 、 and noise Input to A reversible hidden block, calculated by the inverse mapping function , and as follows:

[0056] ;

[0057] ;

[0058] ;

[0059] During the reverse recovery process, reversible hidden blocks, , whose input is Output of each block 、 and , after the inverse transformation in this stage, the output is updated to:

[0060] ;

[0061] ;

[0062] ;

[0063] After the reverse processing of the first reversible hidden block, the system will obtain the initial recovery variable 、 and , these variables correspond to the final state output of the reverse recovery process:

[0064] ;

[0065] ;

[0066] ;

[0067] and It is still in the wavelet domain subband state. In order to restore it to the spatial domain, it is necessary to perform frequency-space domain signal reconstruction through the inverse wavelet transform module. The specific process is described as follows:

[0068] ;

[0069] .

[0070] A further technical solution of the present invention is that the MemoryBank module in step S20 ensures steganographic performance and security through a non-repetitive selection mechanism, and step S20 includes:

[0071] When the total number of secret image blocks does not exceed the number of carrier video frames, the MemoryBank module performs a non-repetitive selection process: all secret image blocks generated by block processing of the secret image are merged into a complete secret image block set, and this secret image block set and the carrier video are input into the memory bank module for frame selection. The obtained carrier video frames and secret blocks are input in parallel into the image-video cross-dimensional steganography model VSHI-Net based on reversible neural networks to complete the steganography. Finally, all video frames containing secret image blocks are synchronously generated through a single steganography operation.

[0072] When the total number of secret image blocks exceeds the number of carrier video frames, multiple secret images are directly merged into a unified input set and input into the memory module. The merged secret image and the selected video frames are jointly input into the image-video cross-dimensional steganography model VSHI-Net based on reversible neural network for parallel steganography.

[0073] A further technical solution of the present invention is to use the loss function in the forward hidden stage Enforce constraints on steganographic video Video frame with original cover Maintain high consistency;

[0074] in, represents the L2 loss, Indicates hidden carrier video frame behind the secret image;

[0075] In the backward recovery process, the goal is to recover the secret information from the cover video as much as possible. The backward loss defined for this purpose is as follows:

[0076] ;

[0077] in represents the L2 loss, Indicates the A secret image, Indicates the A restored image;

[0078] In the optimization process of the image-video cross-dimensional steganography model VSHI-Net based on reversible neural network, by jointly minimizing the forward hidden loss function and the backward recovery loss function To implement network training, its objective function is defined as follows:

[0079] ;

[0080] in, and is a hyperparameter used to balance the quality of the steganographic video frame and the restored image.

[0081] To achieve the above objectives, the present invention also proposes an image-video cross-dimensional steganography system, which includes a memory, a processor, and an image-video cross-dimensional steganography program stored on the processor. When the image-video cross-dimensional steganography program is run by the processor, the steps of the method described above are executed.

[0082] To achieve the above objectives, the present invention also proposes a computer-readable storage medium, which stores an image-video cross-dimensional steganography program. When the image-video cross-dimensional steganography program is run by a processor, the steps of the method described above are executed.

[0083] The beneficial effects of the image-video cross-dimensional steganography method, system, and medium of the present invention are:

[0084] Through the collaborative design of the MemoryBank frame selection mechanism and the reversible neural network frequency domain steganography architecture, this invention achieves efficient cross-dimensional embedding and lossless recovery of secret images into carrier videos, significantly improving the carrier utilization, concealment and robustness of cross-modal steganography, and providing a new technical paradigm for secure communication between heterogeneous image and video carriers. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a flowchart of a preferred embodiment of the image-video cross-dimensional steganography method of the invention;

[0086] Figure 2 This is the model architecture diagram of the image-video cross-dimensional steganography model based on reversible neural network;

[0087] Figure 3 This is a schematic diagram of the MemoryBank module structure;

[0088] Figure 4 It is a schematic diagram of the structure of a reversible neural network;

[0089] Figure 5 It is a schematic diagram of the multi-image hiding process.

[0090] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0091] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0092] Traditional video steganography techniques primarily cover three domains: spatial, frequency, and compression. However, these traditional methods, due to their limited hiding capacity and insufficient concealment, are no longer sufficient for current applications. To overcome these limitations, deep learning has been increasingly applied to this field. In this process, reversible neural networks, with their precise and reversible mathematical properties, offer a new approach to information steganography. This technology utilizes a single model to simultaneously achieve data hiding and lossless recovery.

[0093] However, most current steganography research relies on the premise that the secret information and the carrier form are identical, which limits selectivity and flexibility in practical applications. Furthermore, the few studies focused on cross-dimensional steganography for images and videos still suffer from poor performance and underutilization of the dimensional differences between images and videos.

[0094] To this end, the present invention proposes an image-video cross-dimensional steganography method based on reversible neural networks, which fully explores and utilizes the high-dimensional redundant information of the video carrier relative to the image, and realizes the cross-dimensional information steganography technology of hiding the secret image in the carrier video.

[0095] like Figure 1 As shown, a preferred embodiment of the image-video cross-dimensional steganography method of the present invention includes the following steps:

[0096] Step S10: In the forward hiding stage, the secret image is pre-processed by the sliding window operation of the image cropping module to divide the secret image into secret image blocks, .

[0097] Step S20: The pre-processed secret image block and the carrier video frame are input into the MemoryBank module, and the pre-processed secret image block and the carrier video frame are filtered based on the feature matching degree. The optimal carrier video frames are used for information hiding and no duplication is ensured between frames.

[0098] Step S30: secret image blocks and the filtered The carrier video frames are input in parallel to the image-video cross-dimensional steganography model VSHI-Net based on reversible hidden neural network for steganography processing, and the video frame output containing the secret image block is generated.

[0099] Step S40: In the reverse recovery phase, the video frame containing the secret image block is used as one of the inputs of the image-video cross-dimensional steganography model VSHI-Net based on a reversible hidden neural network. After network processing, the recovered secret image block is generated, and finally the recovered secret image is obtained by image splicing and reassembly.

[0100] The model framework of the image-video cross-dimensional steganography model VSHI-Net based on reversible hidden neural network involved in the present invention is as follows Figure 2 The present invention relates to an image-video cross-dimensional steganography model based on a reversible hidden neural network, named VSHI-Net, as shown in FIG. Figure 2 As shown in Figure 2, it is specifically divided into the forward hiding stage and the reverse recovery stage. In the forward hiding stage, the secret image is first segmented into secret image blocks The pre-processed secret image blocks and the carrier video frames are input into the MemoryBank module. This module selects the secret image blocks based on feature matching. The optimal carrier video frames are used for information hiding and ensure that there is no repeated selection between frames. The secret image blocks and the selected carrier video frames are fed into the VSHI-Net in parallel for steganographic processing, ultimately generating a video frame output containing the secret image blocks. In the reverse recovery phase, the video frame containing the secret image blocks is used as one of the inputs to the VSHI-Net. After processing through the network, the recovered secret image blocks are generated. Finally, the recovered secret image is obtained through image splicing and reconstructing.

[0101] In this embodiment, in step S10, during the image segmentation process, a sliding window is used to segment the secret image into blocks. Since the size of the secret image blocks is inevitably reduced after cropping, which may cause size mismatch, each cropped secret image block needs to be padded to restore its original size.

[0102] In step S40, during the image stitching stage, the restored secret image blocks need to first remove the filling area, and then be stitched and reassembled according to the original spatial coordinates recorded during the cropping process, and finally a complete secret image is reconstructed.

[0103] In this embodiment, step S20 includes:

[0104] Step S201: In the initialization phase, the secret image block sequence and the carrier video are input into the pre-trained ResNet network for feature extraction to obtain the secret image block feature set. and the carrier video frame feature set .

[0105] Step S202, in the selection stage, uses the cosine similarity measurement method to perform sequential matching calculations on the features of each secret image block and all carrier video frame features, and selects the video frame with the highest matching degree as the steganographic carrier of the image block based on the similarity calculation results.

[0106] In order to achieve accurate matching between the secret image block and the optimal carrier video frame, this embodiment designs a MemoryBank module. Figure 3As shown in the figure, the matching process of this module includes two stages: initialization and selection. Specifically, in the initialization stage ( Figure 3 Left), the secret image block sequence and the carrier video are jointly input into the pre-trained ResNet network for feature extraction, and the secret image block feature sets are obtained respectively. and the carrier video frame feature set In the selection phase ( Figure 3 (right) Using the cosine similarity metric, the features of each secret image block are sequentially matched against the features of all carrier video frames. Based on the similarity calculation results, the video frame with the highest matching score is selected as the steganograft for that image block. It is worth noting that to prevent multiple secret image blocks from using the same video frame, which would degrade steganographic security and performance, this module uses an exclusive selection mechanism to ensure that no carrier video frame is selected repeatedly.

[0107] Please refer to Figure 4 The reversible hiding neural network (IHNN) in step S30 includes three parts: a discrete wavelet transform (DWT) module, a reversible hiding module, and an inverse wavelet transform (IWT) module. The network has a bidirectional reversible property: the forward process realizes secret image hiding, and the reverse process completes secret image recovery.

[0108] Traditional pixel domain steganography methods often result in texture replication artifacts and color distortion, which in turn degrades the quality of the steganographic image. To solve this problem, this embodiment proposes a steganography scheme based on wavelet transform: Figure 4 As shown in Figure 1, the input data is first decomposed into low-frequency and high-frequency subbands using a discrete wavelet transform (DWT) before being processed by the reversible hiding module. By operating in the frequency domain, the DWT and inverse wavelet transform (IWT) can more efficiently embed the secret information into the carrier video frame, effectively avoiding the visual distortion risks common in pixel-domain operations.

[0109] The core advantage of wavelet transform is its perfect reconstruction property, which ensures the accurate recovery of embedded information, thereby reducing information loss and improving steganographic performance. This embodiment uses Haar wavelet basis function, which is an ideal choice due to its computational simplicity and efficiency.

[0110] Wavelet transforms also offer computational advantages. The DWT significantly reduces computational complexity by converting video frames of size (B, C, N, H, W) into wavelet subbands of size (B, 4C, N, H / 2, W / 2) (where B, C, N, H, and W represent the batch size, number of channels, number of video frames, height, and width, respectively). This dimensionality reduction accelerates network training while maintaining embedding quality. It's important to note that the bidirectional symmetry of the wavelet transform fully preserves data integrity for end-to-end network training.

[0111] In this embodiment, in the forward hiding process, the input of the reversible hidden neural network (IHNN) includes three parts: a sequence of cropped secret image blocks , optimally matched carrier video frame and Gaussian noise .

[0112] The step S30 includes:

[0113] Step S301: The clipped secret image block sequence is transformed by a discrete wavelet transform module. , optimally matched carrier video frame Perform high and low frequency wavelet decomposition to generate secret image wavelet subbands respectively and carrier video wavelet subbands .

[0114] Step S302: Generate secret image wavelet subband , carrier video wavelet subband With Gaussian noise Input the reversible hidden module for processing and output the wavelet subband containing the secret image block , wavelet domain information residual and noise residual .

[0115] Step S303: wavelet subband , wavelet domain information residual After reconstruction by the inverse wavelet transform module, the video frame containing the secret image block is output and residual signal .

[0116] Among them, the forward hidden process can be formally expressed as:

[0117] .

[0118] During the reverse recovery process, since the lost information cannot be directly obtained , auxiliary variables need to be introduced Assisted recovery.

[0119] Specifically, step S40 includes:

[0120] Step S401: In the reverse recovery process, random sampling variables are introduced , the carrier video frame containing the secret image block is transformed into With random sampling variables Decomposition into wavelet subbands and , while multiplexing Gaussian noise .

[0121] Step S402: Recover the carrier video wavelet subband through the inversion operation of the reversible concealment module , secret image wavelet subband and noise residual .

[0122] Step S403: transform the carrier video wavelet subband into , secret image wavelet subband Reconstruct into the spatial domain and obtain the restored carrier video frames With secret image block .

[0123] The reverse recovery process can be formally expressed as:

[0124] .

[0125] Among them, the forward mapping function and the inverse mapping function A set of shared network parameters.

[0126] It should be noted that the forward hiding process and the reverse recovery process of IHNN strictly satisfy the mathematical reversibility. This means that the forward mapping function and the inverse mapping function Shared network parameter set. This feature significantly reduces the model's parameter complexity while maintaining bidirectional reversible computing capabilities, demonstrating the structural advantages of reversible neural networks in steganography tasks.

[0127] like Figure 4 As shown, in this embodiment, the reversible hiding module is composed of The input of the first reversible hidden block consists of three parts: the carrier video wavelet subband , secret image wavelet subband and Gaussian noise ,in, and Derived from the carrier video frame With secret image block Frequency domain representation after processing by discrete wavelet transform module.

[0128] ;

[0129] .

[0130] Where, Represents the discrete wavelet transform operator; the output of the first reversible hidden block consists of three parts: the primary wavelet subband containing the secret image block , wavelet domain information residual and noise residual , its mathematical expression is as follows:

[0131] ;

[0132] ;

[0133] .

[0134] Where, Defined as a Sigmoid activation function scaled by a constant factor, (·), (·), (·), (·), (·), (·), (·)and (·) is any nonlinear function that can be learned. reversible hidden blocks ( ), whose input is the wavelet subband containing the secret image block output in the previous stage , information residual and noise residual , after transformation in this stage, the output is updated to:

[0135] ;

[0136] ;

[0137] .

[0138] Assume that the reversible hidden module contains reversible hidden blocks, after the After InvB processing, the system will output the final wavelet subband containing the secret image block , cumulative information residual and noise residual ; These three output quantities correspond to the final state variables of the reversible hidden process:

[0139] .

[0140] It should be noted that and are all subband signals in the wavelet domain, and The encrypted video frames and residual signals converted into the spatial domain need to be reconstructed in the frequency-space domain through the inverse wavelet transform module. Its mathematical expression is:

[0141] ;

[0142] Where, Represents the inverse wavelet transform operator.

[0143] like Figure 4 As shown in the figure, in this embodiment, the information flow direction of the reverse recovery process is opposite to that of the forward process, which is shown as The reversible hidden block It is worth noting that during the reverse recovery process, the system only has the video frames containing the secret image blocks. As input, but lose information and noise residual Unavailable; for this reason, this embodiment introduces an auxiliary variable Alternative , and multiplexed Gaussian noise Alternative , in order to achieve the reversibility of hiding and restoring process, Must Obey the same distribution. This embodiment assumes that both obey Gaussian distribution, that is, .

[0144] The specific implementation process of the reverse recovery process is as follows: Video frame containing secret image block and randomly sampled variables First, convert it to the wavelet domain through the DWT module to generate:

[0145] ;

[0146] .

[0147] Afterwards, 、 and noise Input to A reversible hidden block, calculated by the inverse mapping function , and as follows:

[0148] ;

[0149] ;

[0150] .

[0151] During the reverse recovery process, Reversible hidden blocks , whose input is Output of each block 、 and , after the inverse transformation in this stage, the output is updated to:

[0152] ;

[0153] ;

[0154] .

[0155] After the reverse processing of the first reversible hidden block, the system will obtain the initial recovery variable 、 and , these variables correspond to the final state output of the reverse recovery process:

[0156] ;

[0157] ;

[0158] .

[0159] Special instructions are required. and It is still in the wavelet domain subband state. In order to restore it to the spatial domain, it is necessary to perform frequency-space domain signal reconstruction through the inverse wavelet transform module. The specific process is described as follows:

[0160] ;

[0161] .

[0162] In this embodiment, for the multi-image hiding structure, the MemoryBank module ensures the steganographic performance and security through the non-repetitive selection mechanism. However, in the multi-image steganographic task, when the total number of secret image blocks exceeds the number of available carrier video frames (i.e. ), it will inevitably face the challenge of insufficient carrier video frame resources. and Differentiated steganographic strategies are designed for two typical scenarios (e.g. Figure 5 As shown in Figure 3), an adaptive mechanism is used to achieve optimal information hiding under the constraints of system security.

[0163] Specifically, in this embodiment, step S20 includes:

[0164] When the total number of secret image blocks does not exceed the number of carrier video frames ( ), the MemoryBank module performs a non-repeated selection process:

[0165] like Figure 5 As shown, the secret image 1 and secret image 2, generated by block processing, are combined into a complete secret block set. This set, along with the carrier video, is fed into the memory module for frame selection. The resulting carrier video frames and secret blocks are then fed into the VSHI-Net in parallel to complete the steganography. Ultimately, all video frames containing the secret image blocks are synchronously generated through a single steganography operation. Notably, in this mode, all secret blocks and selected video frames are processed in parallel, improving computational efficiency.

[0166] When the total number of secret image blocks exceeds the number of carrier video frames ( ), traditional blocking strategies lead to the reuse of carrier frames, which in turn degrades steganographic performance and security. To address this, this embodiment innovatively employs a non-blocking steganographic scheme: This eliminates the blocking of secret images and directly merges multiple secret images into a unified input set for the memory module. This merged secret image and selected video frames are then fed into the VSHI-Net for parallel steganalysis. This strategy not only avoids the carrier frame reuse issue, but also eliminates the need for image splicing during the recovery phase, as image blocking is not performed during the stegoscopy phase.

[0167] In this embodiment, the loss function is used to constrain two parts: forward hiding and backward recovery. Forward hiding aims to embed a series of secret image blocks into the selected cover video frame. The generated stego video Two conditions should be met: the existence of the secret image block cannot be detected, and it needs to be consistent with the selected cover video frame. As similar as possible. Therefore, this embodiment uses the loss function Enforce constraints on steganographic video Video frame with original cover Maintain high consistency.

[0168] Specifically, in this embodiment, the loss function is used in the forward hidden stage. Enforce constraints on steganographic video Video frame with original cover Maintain high consistency.

[0169] .

[0170] in, represents the L2 loss, Indicates hidden The carrier video frame after the secret image.

[0171] In the backward recovery process, the goal of this embodiment is to recover the secret information from the cover video as much as possible. For this purpose, the backward loss is defined as as follows:

[0172] .

[0173] in represents the L2 loss, Indicates the A secret image, Indicates the A restored image.

[0174] In the optimization process of the image-video cross-dimensional steganography model VSHI-Net based on the reversible neural network, this embodiment minimizes the forward hidden loss function and the backward recovery loss function To implement network training, its objective function is defined as follows:

[0175] .

[0176] in, and is a hyperparameter used to balance the quality of the steganographic video frame and the restored image.

[0177] The beneficial effects of the image-video cross-dimensional steganography method of the present invention are:

[0178] Through the collaborative design of the MemoryBank frame selection mechanism and the reversible neural network frequency domain steganography architecture, this invention achieves efficient cross-dimensional embedding and lossless recovery of secret images into carrier videos, significantly improving the carrier utilization, concealment and robustness of cross-modal steganography, and providing a new technical paradigm for secure communication between heterogeneous image and video carriers.

[0179] To achieve the above objectives, the present invention also proposes an image-video cross-dimensional steganography system, which includes a memory, a processor, and an image-video cross-dimensional steganography program stored on the processor. When the image-video cross-dimensional steganography program is run by the processor, the steps of the method described above are executed, which will not be repeated here.

[0180] To achieve the above objectives, the present invention also proposes a computer-readable storage medium, which stores an image-video cross-dimensional steganography program. When the image-video cross-dimensional steganography program is run by a processor, the steps of the method described above are executed, which will not be repeated here.

[0181] The above description is only a preferred embodiment of the present invention and does not limit the scope of the invention. Any equivalent structure or process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the scope of protection of the present invention.

Claims

1. An image-video cross-dimensional steganography method, characterized in that: The method comprises the following steps: Step S10: In the forward hiding stage, the secret image is pre-processed by the sliding window operation of the image cropping module to divide the secret image into secret image blocks; Step S20: The pre-processed secret image block and the carrier video frame are input into the MemoryBank module, and the pre-processed secret image block and the carrier video frame are filtered based on the feature matching degree. The optimal carrier video frames are used for information hiding and ensure that there is no repeated selection between frames; Step S30: secret image blocks and the filtered The carrier video frames are input in parallel to the image-video cross-dimensional steganography model VSHI-Net based on the reversible hidden neural network for steganography processing, and the video frame output containing the secret image block is generated; Step S40: In the reverse recovery phase, the video frame containing the secret image block is used as one of the inputs of the image-video cross-dimensional steganography model VSHI-Net based on a reversible hidden neural network. After network processing, the recovered secret image block is generated, and finally the recovered secret image is obtained by image splicing and recombination.

2. The image-video cross-dimensional steganography method according to claim 1, characterized in that: In step S10, a sliding window is used to divide the secret image into blocks during the image cutting process. Since the size of the secret image blocks will inevitably decrease after cropping, a size mismatch problem will occur. Therefore, each cropped secret image block needs to be padded to restore the original size. In step S40, during the image stitching stage, the restored secret image blocks need to first remove the padded area, and then be spliced ​​and reassembled according to the original spatial coordinates recorded during the cropping process, and finally reconstruct the complete secret image.

3. The image-video cross-dimensional steganography method according to claim 2, characterized in that: The step S20 includes: Step S201: In the initialization phase, the secret image block sequence and the carrier video are input into the pre-trained ResNet network for feature extraction to obtain the secret image block feature set. and the carrier video frame feature set ; Step S202, in the selection stage, uses the cosine similarity measurement method to perform sequential matching calculations on the features of each secret image block and all carrier video frame features, and selects the video frame with the highest matching degree as the steganographic carrier of the image block based on the similarity calculation results.

4. The image-video cross-dimensional steganography method according to claim 3, characterized in that: The reversible hidden neural network in step S30 includes three parts: discrete wavelet transform module, reversible hidden module and inverse wavelet transform module. In the forward hiding process, the input of the reversible hidden neural network includes three parts: the cropped secret image block sequence , optimally matched carrier video frame and Gaussian noise ; The step S30 includes: Step S301: The clipped secret image block sequence is transformed by a discrete wavelet transform module. , optimally matched carrier video frame Perform high and low frequency wavelet decomposition to generate secret image wavelet subbands respectively and carrier video wavelet subbands ; Step S302: Generate secret image wavelet subband , carrier video wavelet subband With Gaussian noise Input the reversible hidden module for processing and output the wavelet subband containing the secret image block , wavelet domain information residual and noise residual ; Step S303: wavelet subband , wavelet domain information residual After reconstruction by the inverse wavelet transform module, the video frame containing the secret image block is output and residual signal ; Among them, the forward hidden process can be formally expressed as: ； The step S40 includes: introducing random sampling variables in the reverse recovery process , the carrier video frame containing the secret image block is transformed into With random sampling variables Decomposition into wavelet subbands and , while multiplexing Gaussian noise ; Through the inversion operation of the reversible hidden module, the carrier video wavelet subband is restored , secret image wavelet subband and noise residual ; The carrier video wavelet subband is transformed into , secret image wavelet subband Reconstruct into the spatial domain and obtain the restored carrier video frames With secret image block ; The reverse recovery process can be formally expressed as: ； Forward mapping function and the inverse mapping function A set of shared network parameters.

5. The image-video cross-dimensional steganography method according to claim 4, characterized in that: The reversible hiding module is composed of The first reversible hidden block input consists of three parts: the carrier video wavelet subband , secret image wavelet subband and Gaussian noise ,in, and Derived from the carrier video frame Frequency domain representation of the secret image block after processing by discrete wavelet transform module ; ； ； Where, Represents the discrete wavelet transform operator; the output of the first reversible hidden block consists of three parts: the primary wavelet subband containing the secret image block , wavelet domain information residual and noise residual , its mathematical expression is as follows: ； ； ； Where, Defined as a sigmoid activation function scaled by a constant factor, (·), (·), (·), (·), (·), (·), (·)and (·) is any nonlinear function that can be learned. Reversible hidden blocks ( ), whose input is the wavelet subband containing the secret image block output by the previous stage , information residual and noise residual , after transformation in this stage, the output is updated to: ； ； ； Assume that the reversible hidden module contains reversible hidden blocks, after the After processing the reversible hidden blocks, the system will output the final wavelet subband containing the secret image block , cumulative information residual and noise residual ; These three output quantities correspond to the final state variables of the reversible hidden process: ； and are all subband signals in the wavelet domain, and The encrypted video frames and residual signals converted into the spatial domain need to be reconstructed in the frequency-space domain through the inverse wavelet transform module. The mathematical expression is: ； ； Where, Represents the inverse wavelet transform operator.

6. The image-video cross-dimensional steganography method according to claim 5, characterized in that: The direction of information flow in the reverse recovery process is opposite to that in the forward process, which is manifested as The reversible hidden block Reversible hidden block is transmitted backward; in the reverse recovery process, the system only has the video frame containing the secret image block As input, but loses information and noise residual Unavailable; for this reason, an auxiliary variable is introduced Alternative , and multiplexed Gaussian noise Alternative , in order to achieve the reversibility of hiding and restoring process, Must obey the same distribution; assuming that both obey Gaussian distribution, that is, ; The specific implementation process of the reverse recovery process is as follows: Video frame containing secret image block and randomly sampled variables First, convert it to the wavelet domain through the discrete wavelet transform module to generate: ； Afterwards, 、 and noise Input to A reversible hidden block, calculated by the inverse mapping function , and as follows: ； ； ； During the reverse recovery process, reversible hidden blocks, , whose input is Output of each block 、 and , after the inverse transformation in this stage, the output is updated to: ； ； ； After the reverse processing of the first reversible hidden block, the system will obtain the initial recovery variable 、 and , these variables correspond to the final state output of the reverse recovery process: ； ； ； and It is still in the wavelet domain subband state. In order to restore it to the spatial domain, it is necessary to perform frequency-space domain signal reconstruction through the inverse wavelet transform module. The specific process is described as follows: ； 。 7. The image-video cross-dimensional steganography method according to claim 6, characterized in that: In step S20, the MemoryBank module ensures steganographic performance and security through a non-repetitive selection mechanism. Step S20 includes: When the total number of secret image blocks does not exceed the number of carrier video frames, the MemoryBank module performs a non-repetitive selection process: all secret image blocks generated by block processing of the secret image are merged into a complete secret image block set, and this secret image block set and the carrier video are input into the memory bank module for frame selection. The obtained carrier video frames and secret blocks are input in parallel into the image-video cross-dimensional steganography model VSHI-Net based on reversible neural networks to complete the steganography. Finally, all video frames containing secret image blocks are synchronously generated through a single steganography operation. When the total number of secret image blocks exceeds the number of carrier video frames, multiple secret images are directly merged into a unified input set and input into the memory module. The merged secret image and the selected video frames are jointly input into the image-video cross-dimensional steganography model VSHI-Net based on reversible neural network for parallel steganography.

8. The image-video cross-dimensional steganography method according to claim 7, characterized in that: In the forward hidden stage, the loss function Enforce constraints on steganographic video Video frame with original cover Maintain high consistency; in, represents the L2 loss, Indicates hidden carrier video frame behind the secret image; In the backward recovery process, the goal is to recover the secret information from the cover video as much as possible. The backward loss defined for this purpose is as follows: ； in represents the L2 loss, Indicates the A secret image, Indicates the A restored image; In the optimization process of the image-video cross-dimensional steganography model VSHI-Net based on reversible neural network, by jointly minimizing the forward hidden loss function and the backward recovery loss function To implement network training, its objective function is defined as follows: ； in, and is a hyperparameter used to balance the quality of the steganographic video frame and the restored image.

9. An image-video cross-dimensional steganography system, characterized in that: The system includes a memory, a processor, and an image-video cross-dimensional steganography program stored on the processor. When the image-video cross-dimensional steganography program is run by the processor, the steps of the method according to any one of claims 1 to 7 are executed.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores an image-video cross-dimensional steganography program, and when the image-video cross-dimensional steganography program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are executed.

Images