Video steganography method and system based on hybrid scale reversible attention

Abstract

This invention belongs to the field of information hiding technology and discloses a video steganography method and system based on hybrid-scale reversible attention. First, the cover and secret video frames are mapped to the frequency domain via discrete wavelet transform, providing a foundation for reversible information fusion. Then, a pre-trained end-to-end fully reversible video steganography neural network is used to complete multi-level reversible transformations through multiple cascaded multi-scale hybrid reversible blocks, fundamentally solving the problem of irreversible information loss in traditional CNN steganography methods, while simultaneously achieving effective fusion of multi-scale spatiotemporal features. The transformation results are input into a dual-attention redundancy refinement module to generate residual information. In the recovery stage, after the steganographic video frame undergoes discrete wavelet transform, the inverse transform of each level of reversible block is executed using network shared parameters, and the original features are separated and restored by combining the residual information. This fully utilizes video visual redundancy to compensate for information loss, and finally, the original video frame is recovered via inverse discrete wavelet transform, solving the technical problems existing in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of information hiding technology, specifically relating to a video steganography method and system based on hybrid scale reversible attention. Background Technology

[0002] Video steganography is a technique that embeds secret information into digital video carriers to achieve covert communication or copyright protection. Traditional video steganography methods mainly focus on modifying pixel values, transform domain coefficients (such as DCT and DWT), or motion vectors of video frames. In recent years, with the development of deep learning, steganography methods based on convolutional neural networks (CNNs) have become a research hotspot due to their powerful feature extraction and adaptive embedding capabilities. These methods typically use an encoder-decoder architecture to embed secret information into the carrier video to generate a steganized video, and then extract the secret information at the receiving end using a decoder. However, to ensure the visual quality of the steganized video and the accuracy of secret information extraction, model design is crucial.

[0003] Despite some progress in existing technologies, the following significant drawbacks still exist: 1. Information loss and irreversibility: Most CNN-based steganalysis networks use downsampling (such as pooling and strided convolution) and non-linear activation functions. These operations are inherently irreversible, leading to the loss of original information during embedding and extraction, which affects the quality of the reconstructed video and the accuracy of extracting secret information.

[0004] 2. Insufficient utilization of multi-scale features: Video content has rich multi-scale spatiotemporal characteristics. Existing methods often fail to effectively integrate contextual information at different scales, resulting in the embedded secret information being easily destroyed or difficult to accurately recover in complex textures or dynamic scenes.

[0005] 3. Trade-off between efficiency and effectiveness of attention mechanisms: While attention mechanisms can help models focus on important regions, global attention is computationally expensive, while local attention may ignore long-distance dependencies. Designing an attention module that can capture global context while preserving local details and is computationally efficient remains a challenge in the field of video steganography.

[0006] 4. Inadequate handling of redundant information: There is a large amount of visual redundancy between and within video frames. Existing methods fail to fully utilize this redundancy to provide a more secure "hiding space" for secret information, thus limiting steganography capacity and robustness.

[0007] Therefore, existing video steganography methods generally suffer from irreversible information loss, insufficient fusion of multi-scale spatiotemporal features, difficulty in balancing the efficiency and effectiveness of attention mechanisms, and insufficient utilization of visual redundancy. This results in poor quality of reconstructed videos, low accuracy in extracting secret information, and limited steganography capacity and robustness against attacks. Summary of the Invention

[0008] The purpose of this invention is to solve the problems of irreversible information loss, insufficient multi-scale fusion, and improper use of redundancy in the prior art, and to provide a video steganography method and system based on hybrid scale reversible attention.

[0009] To achieve the above objectives, the present invention employs the following technical solution: This invention proposes a video steganography method based on hybrid scale reversible attention, comprising the following steps: Obtain the cover video frame and the secret video frame to be hidden, and perform discrete wavelet transform on the two types of video frames respectively to map them to the frequency domain to obtain the cover frequency domain features and the secret frequency domain features. The obtained cover frequency domain features and secret frequency domain features are input into a pre-trained end-to-end fully reversible video steganalysis neural network. After passing through multiple cascaded multi-scale hybrid reversible blocks, a multi-level reversible transformation is completed. The transformation results are input into a dual-attention redundancy refinement module to generate residual information. The transformation result is reconstructed back into the spatial domain by inverse discrete wavelet transform to generate steganalytic video frames. Based on the steganalytic video frames and residual information, the original video features are separated and restored, and then the original video frames are restored at the pixel level by inverse discrete wavelet transform, thus realizing video steganalysis based on hybrid scale reversible attention.

[0010] Preferably, the multi-level reversible transformation is completed through multiple cascaded multi-scale hybrid reversible blocks, specifically as follows: In each multi-scale hybrid reversible block, the SimpleGate structure is first used to replace the traditional ReLU activation function, and nonlinearity is introduced by grouping channels and performing element-wise multiplication. Then, the features are extracted into multi-scale features through parallel large-kernel convolution and standard convolution paths, and combined with dilated convolution to capture contextual information of different receptive fields. The channel attention submodule is simplified to model the dependencies between channels and adaptively recalibrate the importance of each channel feature. Finally, the deep and shallow features are aggregated through dense connections.

[0011] Preferably, in the forward hiding process, each multi-scale hybrid reversible block receives the cover features and secret features of the previous level in each transformation. Through a learnable nonlinear function, the features are additively coupled and affinely transformed to achieve the gradual fusion of secret information and cover information.

[0012] Preferably, the step of inputting the transformation result into the dual-attention redundancy refinement module to generate residual information specifically involves: Multi-level reversible transformation is completed through multiple cascaded multi-scale hybrid reversible blocks to achieve the gradual fusion of secret information and cover information, and the fusion result is recorded as the transformation result; The transformation result is fed into a dual-attention multi-scale fusion submodule, which adopts a dual-branch parallel architecture. One branch uses a non-local sparse attention mechanism based on locality-sensitive hashing to establish long-range spatial dependencies within video frames and temporal correlations between frames. The other branch extracts and enhances features of local regions through convolution and channel shuffling operations. The output features of the two branches are adaptively fused with the original input features to form a redundant feature representation. The redundant feature representation is then input into a redundant prediction refinement submodule for extraction to generate residual information used to compensate for information loss.

[0013] Preferably, the step of inputting the redundant feature representation into the redundancy prediction refinement submodule for extraction specifically involves: A feedforward network with dual-branch deep convolutions is used to perform multi-scale filtering on redundant feature representations of the input; information-rich channels are adaptively enhanced and redundant channels are suppressed through a channel attention mechanism; and features are progressively optimized and purified by stacking multiple residual blocks to eliminate irrelevant redundant information while strengthening effective features related to video motion patterns and texture structures.

[0014] Preferably, the pre-trained end-to-end fully reversible video steganalysis neural network specifically comprises: End-to-end joint optimization is performed using a multi-objective loss function, which is a weighted sum of a loss module and an enhancement loss module. The loss module includes forward fitting loss, inverse reconstruction loss, and secret recovery loss; the enhancement loss module includes sparse loss, time-aware loss, and time-series gradient smoothing loss.

[0015] Preferably, the forward fitting loss is used to enforce the visual consistency between the stegana video and the cover video, the inverse reconstruction loss is used to enforce the consistency between the recovered cover video and the original cover video, and the secret recovery loss is used to enforce the matching between the recovered secret video and the original secret video. Sparse loss is used to constrain the magnitude and distribution of modifications to the original video features during the embedding process, temporal-aware loss is used to utilize the features of the pre-trained spatiotemporal network to maintain inter-frame consistency, and temporal gradient smoothing loss is used to constrain the smooth changes in differences between adjacent frames.

[0016] This invention proposes a video steganography system based on hybrid scale reversible attention, comprising: The preprocessing module is used to acquire the cover video frame and the secret video frame to be hidden, and to perform discrete wavelet transform on the two types of video frames respectively to map them to the frequency domain, so as to obtain the cover frequency domain features and the secret frequency domain features. The steganography processing module is used to input the cover frequency domain features and secret frequency domain features into a pre-trained end-to-end fully reversible video steganography neural network, and complete multi-level reversible transformation through multiple cascaded multi-scale hybrid reversible blocks. The transformation result is then input into a dual attention redundancy refinement module to generate residual information. The generation and restoration module is used to reconstruct the transformation result back to the spatial domain through inverse discrete wavelet transform to generate a steganalytic video frame; based on the steganalytic video frame and residual information, the original video features are separated and restored, and then the original video frame at the pixel level is restored through inverse discrete wavelet transform, thus realizing video steganalysis based on hybrid scale reversible attention.

[0017] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the video steganography method based on hybrid scale reversible attention.

[0018] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the video steganography method based on hybrid scale reversible attention.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a video steganography method based on hybrid-scale reversible attention. First, the cover and secret video frames are mapped from the spatial domain to the frequency domain using discrete wavelet transform to obtain corresponding frequency domain features. This step provides a suitable processing space for subsequent reversible information fusion and decomposes the video information into different frequency sub-bands, laying the foundation for multi-scale feature processing. Second, a pre-trained end-to-end fully reversible video steganography neural network is employed. Through multiple cascaded multi-scale hybrid reversible blocks, a rigorous mathematical reversible transformation is performed, fundamentally avoiding the information loss problem caused by irreversible operations such as downsampling and nonlinear activation in traditional CNN steganography methods. Furthermore, the design of the multi-scale hybrid reversible blocks effectively fuses spatiotemporal features at different scales, addressing the deficiency of insufficient feature utilization in complex scenarios. Based on this, the reversible transformation results are input into a dual-attention redundancy refinement module to generate residual information, actively modeling and utilizing the inherent visual redundancy of the video. Finally, the reversible transformation result is used to generate steganalytic video frames through inverse discrete wavelet transform. In the recovery stage, based on the steganalytic video frames and the generated residual information, the inverse transformation of each level of reversible blocks is performed inversely using network shared parameters. The original video features are separated and restored, and then the original video frames are restored at the pixel level through inverse transformation. The residual information is used to compensate for potential information loss, and finally the lossless recovery of the original cover and the secret video is achieved, solving the problems of irreversible information loss, insufficient multi-scale fusion and improper use of redundancy in the existing technology.

[0020] This invention proposes a video steganography system based on hybrid scale reversible attention. By dividing the system into a preprocessing module, a steganography processing module, and a generation and restoration module, it achieves video steganography based on hybrid scale reversible attention. The modular approach ensures that each module is independent, facilitating unified management of all modules. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the video steganography method based on hybrid scale reversible attention of the present invention.

[0023] Figure 2 This is a detailed flowchart of the steganography method of the present invention.

[0024] Figure 3 This is a diagram of the internal structure of the multi-scale hybrid reversible block of the present invention.

[0025] Figure 4 This is a diagram of the multi-scale hybrid reversible block structure of the present invention.

[0026] Figure 5 This is a structural diagram of the dual attention redundancy refinement module of the present invention.

[0027] Figure 6 This is a diagram of the video steganography system based on hybrid scale reversible attention according to the present invention.

[0028] Figure 7 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0030] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0031] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0032] The present invention will now be described in further detail with reference to the accompanying drawings: The definitions are as follows: 1) Invertible Neural Network: A special network structure whose forward computation process is mathematically reversible. Given the output and network parameters, a unique input can be accurately reconstructed. This is the basis for the lossless embedding / extraction guaranteed in this invention.

[0033] (2) Discrete Wavelet Transform (DWT): A signal processing technique that decomposes an image / video into subbands of different frequencies (such as low-frequency approximations and high-frequency details). In this invention, DWT is used as a preprocessing step to transform the data into a frequency domain more suitable for embedding.

[0034] (3) Visual Redundancy: refers to information regions in a video that are not easily perceived by the human eye or contribute little to visual perception. Embedding secret information in these regions can improve the invisibility of steganography.

[0035] Based on the technical problems existing in the background technology, this invention proposes a video steganography method based on hybrid-scale reversible attention. Employing a symmetric reversible architecture, it achieves the cover video's cover-up with the concealment of one or more secret videos, supporting lossless recovery. The core of this invention lies in effectively solving the problems of information loss, insufficient spatiotemporal redundancy modeling, and degraded recovery quality in existing reversible neural networks for video steganography through multi-scale hybrid reversible blocks and dual-attention redundant refinement modules (DARRM). Figure 1 As shown, the video steganography method based on hybrid scale reversible attention includes the following steps: S1. Obtain the cover video frame and the secret video frame to be hidden. Perform discrete wavelet transform on the two types of video frames respectively to map them to the frequency domain to obtain the cover frequency domain features and the secret frequency domain features. S2. Input the obtained cover frequency domain features and secret frequency domain features into the pre-trained end-to-end fully reversible video steganalysis neural network. After multiple cascaded multi-scale hybrid reversible blocks, multi-level reversible transformation is completed. The transformation results are input into the dual attention redundancy refinement module to generate residual information. The multi-level reversible transformation, achieved through multiple cascaded multi-scale hybrid reversible blocks, is specifically as follows: In each multi-scale hybrid reversible block, the SimpleGate structure is first used to replace the traditional ReLU activation function, and nonlinearity is introduced by grouping channels and performing element-wise multiplication. Then, the features are extracted into multi-scale features through parallel large-kernel convolution and standard convolution paths, and combined with dilated convolution to capture contextual information of different receptive fields. The channel attention submodule is simplified to model the dependencies between channels and adaptively recalibrate the importance of each channel feature. Finally, the deep and shallow features are aggregated through dense connections.

[0036] In each multi-scale hybrid reversible block, during the forward hiding process, each transformation level receives the cover features and secret features from the previous level. Through a learnable nonlinear function, additive coupling and affine transformation are performed on the features to achieve the gradual fusion of secret information and cover information.

[0037] The step of inputting the transformation result into the dual-attention redundancy refinement module to generate residual information is as follows: Multi-level reversible transformation is completed through multiple cascaded multi-scale hybrid reversible blocks to achieve the gradual fusion of secret information and cover information, and the fusion result is recorded as the transformation result; The transformation result is fed into a dual-attention multi-scale fusion submodule, which adopts a dual-branch parallel architecture. One branch uses a non-local sparse attention mechanism based on locality-sensitive hashing to establish long-range spatial dependencies within video frames and temporal correlations between frames. The other branch extracts and enhances features of local regions through convolution and channel shuffling operations. The output features of the two branches are adaptively fused with the original input features to form a redundant feature representation. The redundant feature representation is then input into a redundant prediction refinement submodule for extraction to generate residual information used to compensate for information loss.

[0038] The process of inputting the redundant feature representation into the redundancy prediction refinement submodule for extraction is specifically as follows: A feedforward network with dual-branch deep convolutions is used to perform multi-scale filtering on redundant feature representations of the input; information-rich channels are adaptively enhanced and redundant channels are suppressed through a channel attention mechanism; and features are progressively optimized and purified by stacking multiple residual blocks to eliminate irrelevant redundant information while strengthening effective features related to video motion patterns and texture structures.

[0039] The pre-trained end-to-end fully reversible video steganalysis neural network is specifically as follows: End-to-end joint optimization is performed using a multi-objective loss function, which is a weighted sum of a loss module and an enhancement loss module. The loss module includes forward fitting loss, inverse reconstruction loss, and secret recovery loss; the enhancement loss module includes sparse loss, time-aware loss, and time-series gradient smoothing loss.

[0040] Forward fitting loss is used to enforce visual consistency between the stegana video and the cover video, inverse reconstruction loss is used to enforce consistency between the recovered cover video and the original cover video, and secret recovery loss is used to enforce matching between the recovered secret video and the original secret video. Sparse loss is used to constrain the magnitude and distribution of modifications to the original video features during the embedding process, temporal-aware loss is used to utilize the features of the pre-trained spatiotemporal network to maintain inter-frame consistency, and temporal gradient smoothing loss is used to constrain the smooth changes in differences between adjacent frames.

[0041] S3. The transformation result is reconstructed back into the spatial domain by inverse discrete wavelet transform to generate a steganalytic video frame. Based on the steganalytic video frame and residual information, the original video features are separated and restored, and then the original video frame at the pixel level is restored by inverse discrete wavelet transform, thus realizing video steganalysis based on hybrid scale reversible attention.

[0042] The following is in conjunction with the appendix Figure 2 To be continued Figure 5 The method is described in detail below: Compared to Figure 2 The present invention represents a fundamental improvement over the earlier reversible network methods shown in (a) above. Early methods used only simple reversible blocks for information transfer, which easily discarded redundant information crucial for recovery during complex transformations, leading to a decline in recovery quality. The present invention, however, enhances feature representation and preservation capabilities by introducing multi-scale hybrid reversible blocks and innovatively adds a DARRM module to actively model, optimize, and utilize spatiotemporal redundancy. Figure 2 In (b) of this invention, the framework explicitly processes and retains "modular redundant information" in the hidden path through DAMSF and RPRM modules. This information is reused in the recovery path, thereby achieving a leap from "low-quality recovery" to "high-quality recovery." Experiments show that this scheme significantly outperforms existing state-of-the-art methods in terms of steganalytic video quality, secret recovery accuracy, resistance to steganalysis, and robustness against various attacks, while maintaining high embedding capacity.

[0043] like Figure 3 As shown, the overall architecture of this invention presents a strictly symmetrical forward hiding and backward recovery data flow. Its input includes a cover video frame. And N secret video frames to be hidden The video data is first mapped to the frequency domain using Discrete Wavelet Transform (DWT). This step converts spatial pixel information into frequency subbands, laying the foundation for subsequent reversible information fusion and embedding in the frequency domain. In the frequency domain, the data stream undergoes multi-level processing through a series of cascaded multi-scale hybrid reversible blocks. The core function of each multi-scale hybrid reversible block is to perform a parameter-shared reversible transformation, which maintains the mathematically rigorous reversibility of the transformation process while integrating secret information into cover features, thus laying the foundation for lossless recovery. Specifically, in the forward hiding process, each transformation level receives cover features and secret features from the previous level. Through a learnable nonlinear function, additive coupling and affine transformations are applied to the features to achieve progressive information fusion. After multiple transformations, the resulting features... Figure 1 On the one hand, the residual information z is generated by the dual attention redundancy refinement module DARRM to compensate for information loss; on the other hand, the spatial domain is reconstructed through inverse discrete wavelet transform IDWT, and finally, a steganalytic video frame that is visually highly similar to the original cover video is output. The backward recovery process is the exact inverse operation of the forward process. Starting with the stegated video frame, after undergoing the same DWT transformation, it uses network-shared parameters to perform the inverse transformation of each level of the multi-scale hybrid reversible block in the reverse order of the hiding process, and combines this with the residual information saved or generated during the forward process. The original cover video features and the features of each secret video were gradually separated and restored, and finally the pixel-level video frames were recovered again through IDWT.

[0044] Multiscale hybrid reversible blocks, whose internal structure is as follows Figure 4As shown, this module is not a single reversible layer, but a carefully designed composite unit designed to enhance the model's ability to represent the global context and local details of the video while strictly maintaining reversibility. The module first uses a SimpleGate structure to replace the traditional ReLU activation function. This structure introduces non-linearity by grouping channels and performing element-wise multiplication, effectively avoiding information loss caused by activation function saturation or truncation, thus facilitating the preservation of subtle video details. Subsequently, features are extracted using parallel large-kernel convolutions and standard convolutional paths to extract multi-scale features, combined with dilated convolutions to capture contextual information from different receptive fields. The module also integrates a Simplified Channel Attention (SCA) submodule, which models the dependencies between channels using global average pooling and 1x1 convolutions, adaptively recalibrating the importance of each channel's features to enhance the extraction of key information. The entire multi-scale hybrid reversible block further aggregates features from shallow and deep layers through dense connections (DenseBlock), ensuring effective gradient propagation and mitigating the gradient vanishing problem. All of these operations are designed with reversibility in mind, so that the entire module can retain and merge information to the maximum extent during forward propagation, and accurately trace back to the original input during reverse recovery.

[0045] To more accurately model and utilize the inherent spatiotemporal redundancy information in videos, this invention proposes a dual-attention redundancy refinement module (DARRM). This module consists of a dual-attention multi-scale fusion submodule (DAMSF) and a redundancy prediction refinement submodule (RPRM) working collaboratively. Figure 5As shown, the DAMSF submodule employs a dual-branch parallel architecture. One branch uses a nonlocal sparse attention mechanism based on Locality Sensitive Hashing (LSH), which can establish long-range spatial dependencies within video frames and temporal relationships between frames with low computational complexity, thereby capturing global spatiotemporal context information. The other branch focuses on local details, extracting and enhancing features of local regions through convolution and channel shuffling operations. The output features of the two branches are adaptively fused with the original input features at multiple scales to form redundant feature representations rich in global context and local details. Subsequently, these features are fed into the RPRM submodule for further refinement. RPRM first uses a feedforward network FFN with dual-branch depthwise convolutions to perform multi-scale filtering on the input, and adaptively enhances information-rich channels and suppresses redundant channels through a channel attention mechanism. Then, by stacking multiple residual blocks, the features are progressively optimized and purified, gradually eliminating irrelevant redundant information while strengthening effective features related to video motion patterns and texture structures. The DARRM module plays the role of "redundancy sensing and scheduler" in the whole system. It not only guides information to be embedded into visually insensitive redundant areas during the forward hiding process, but also provides additional clues through residual information z during the backward recovery process, effectively compensating for information loss that may be caused by transformation or attack, thereby significantly improving the fidelity of recovery and the robustness of the system.

[0046] This invention employs a multi-objective loss function for end-to-end joint optimization of the entire network. This loss function is a weighted average of a core loss module and an enhancement loss module. The core loss module enforces three fundamental objectives: visual consistency between the steganalysis and the cover video (forward fitting loss), strict consistency between the recovered cover video and the original cover video (inverse reconstruction loss), and precise matching between the recovered secret video and the original secret video (secret recovery loss). The enhancement loss module introduces deeper constraints, including sparse loss to encourage minimal and sparse embedding modifications, temporal-aware loss to maintain inter-frame consistency using pre-trained spatiotemporal network features, and temporal gradient smoothing loss to constrain the smooth changes in differences between adjacent frames to maintain natural and fluid motion. These loss terms work together to guide the network to achieve high-capacity information hiding while ensuring excellent visual imperceptibility, strict lossless reversibility, and good temporal coherence. Therefore, this invention proposes an end-to-end fully reversible video steganography neural network architecture (VSN), which is used in the forward direction to embed secret information to generate a steganographic video, and in the reverse direction to losslessly separate the original carrier video and secret information from the steganographic video. A hybrid-scale reversible attention module integrated into the above reversible network is itself reversibly designed to ensure no information loss is introduced during the forward / backward propagation of the network. It includes a reversible hybrid-scale perceptual block (MixInvBlock), which achieves effective extraction and fusion of multi-scale local features under reversible constraints through parallel multi-scale depthwise separable convolution and gating mechanisms. It also includes a dual-attention redundancy refinement module (DARRM), which combines LSH-based sparse nonlocal attention (for capturing global spatiotemporal dependencies) and convolutional attention (for preserving local details), and utilizes the fused features to predict and utilize video visual redundancy for secret information embedding. Compared with existing technologies, this invention has the following outstanding advantages: 1) Significantly reduces information loss in complex scenes: By introducing a multi-scale hybrid reversible block MixInvBlock, this invention achieves effective fusion of global and local features under strict reversibility constraints. This design ensures high fidelity of information during the forward (embedding) and backward (extraction) processes of the network, especially in video scenes with complex textures and rapid dynamics, effectively maintaining the reconstruction quality of the carrier video and secret information.

[0047] 2) Precise Modeling and Utilization of Spatiotemporal Redundancy: The Dual Attention Multi-Scale Fusion Module (DAMSF) innovatively combines local convolutional attention with non-local sparse attention mechanisms, enabling it to simultaneously capture fine local details and long-range spatiotemporal dependencies. This dual attention mechanism allows the network to more accurately identify and quantify visual redundancy in videos, providing ideal embedding locations for secret information.

[0048] 3) Achieving an optimal balance between capacity, invisibility, and security: The Redundancy Prediction Refinement (RPRM) module, based on DAMSF's accurate modeling of redundancy, adaptively optimizes the feature representation used for embedding. By intelligently embedding secret information into predicted high-redundancy regions, this invention significantly improves steganography capacity and enhances robustness against potential analytical attacks while ensuring the visual invisibility (high fidelity) of the encrypted video, thus achieving synergistic optimization across the three core metrics.

[0049] Example 2 This invention proposes a video steganography system based on hybrid scale reversible attention, such as... Figure 6 As shown, it includes: The preprocessing module is used to acquire the cover video frame and the secret video frame to be hidden, and to perform discrete wavelet transform on the two types of video frames respectively to map them to the frequency domain, so as to obtain the cover frequency domain features and the secret frequency domain features. The steganography processing module is used to input the cover frequency domain features and secret frequency domain features into a pre-trained end-to-end fully reversible video steganography neural network, and complete multi-level reversible transformation through multiple cascaded multi-scale hybrid reversible blocks. The transformation result is then input into a dual attention redundancy refinement module to generate residual information. The generation and restoration module is used to reconstruct the transformation result back to the spatial domain through inverse discrete wavelet transform to generate a steganalytic video frame; based on the steganalytic video frame and residual information, the original video features are separated and restored, and then the original video frame at the pixel level is restored through inverse discrete wavelet transform, thus realizing video steganalysis based on hybrid scale reversible attention.

[0050] Example 3 Please see Figure 7 As shown, the present invention also provides an electronic device 100 based on a hybrid scale reversible attention video steganography method; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0051] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the video steganography method based on hybrid scale reversible attention described in Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0052] The at least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor. The processor 102 is the control center of the electronic device 100, connecting various parts of the electronic device 100 via various interfaces and lines.

[0053] The memory 101 in the electronic device 100 stores multiple instructions to implement a video steganography method based on hybrid scale reversible attention, and the processor 102 can execute the multiple instructions to achieve the following: Obtain the cover video frame and the secret video frame to be hidden, and perform discrete wavelet transform on the two types of video frames respectively to map them to the frequency domain to obtain the cover frequency domain features and the secret frequency domain features. The obtained cover frequency domain features and secret frequency domain features are input into a pre-trained end-to-end fully reversible video steganalysis neural network. After passing through multiple cascaded multi-scale hybrid reversible blocks, a multi-level reversible transformation is completed. The transformation results are input into a dual-attention redundancy refinement module to generate residual information. The transformation result is reconstructed back into the spatial domain by inverse discrete wavelet transform to generate steganalytic video frames. Based on the steganalytic video frames and residual information, the original video features are separated and restored, and then the original video frames are restored at the pixel level by inverse discrete wavelet transform, thus realizing video steganalysis based on hybrid scale reversible attention.

[0054] Example 4 If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0055] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0058] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A video steganography method based on hybrid scale reversible attention, characterized in that, Includes the following steps: Obtain the cover video frame and the secret video frame to be hidden, and perform discrete wavelet transform on the two types of video frames respectively to map them to the frequency domain to obtain the cover frequency domain features and the secret frequency domain features. The obtained cover frequency domain features and secret frequency domain features are input into a pre-trained end-to-end fully reversible video steganalysis neural network. After passing through multiple cascaded multi-scale hybrid reversible blocks, a multi-level reversible transformation is completed. The transformation results are input into a dual-attention redundancy refinement module to generate residual information. The transformation result is reconstructed back into the spatial domain using inverse discrete wavelet transform to generate steganalytic video frames. Based on the separation and restoration of original video features from stegated video frames and residual information, and then the pixel-level original video frames are recovered through inverse discrete wavelet transform, video steganography based on hybrid scale reversible attention is realized.

2. The video steganography method based on hybrid scale reversible attention according to claim 1, characterized in that, The multi-level reversible transformation, achieved through multiple cascaded multi-scale hybrid reversible blocks, is specifically as follows: In each multi-scale hybrid reversible block, the SimpleGate structure is first used to replace the traditional ReLU activation function, and nonlinearity is introduced by grouping channels and performing element-wise multiplication. Then, the features are extracted into multi-scale features through parallel large-kernel convolution and standard convolution paths, and combined with dilated convolution to capture contextual information of different receptive fields. The channel attention submodule is simplified to model the dependencies between channels and adaptively recalibrate the importance of each channel feature. Finally, the deep and shallow features are aggregated through dense connections.

3. The video steganography method based on hybrid scale reversible attention according to claim 2, characterized in that, In each multi-scale hybrid reversible block, during the forward hiding process, each transformation level receives the cover features and secret features from the previous level. Through a learnable nonlinear function, additive coupling and affine transformation are performed on the features to achieve the gradual fusion of secret information and cover information.

4. The video steganography method based on hybrid scale reversible attention according to claim 1, characterized in that, The step of inputting the transformation result into the dual-attention redundancy refinement module to generate residual information is as follows: Multi-level reversible transformation is completed through multiple cascaded multi-scale hybrid reversible blocks to achieve the gradual fusion of secret information and cover information, and the fusion result is recorded as the transformation result; The transformation result is fed into a dual-attention multi-scale fusion submodule, which adopts a dual-branch parallel architecture. One branch uses a non-local sparse attention mechanism based on locality-sensitive hashing to establish long-range spatial dependencies within video frames and temporal correlations between frames. The other branch extracts and enhances features of local regions through convolution and channel shuffling operations. The output features of the two branches are adaptively fused with the original input features to form a redundant feature representation. The redundant feature representation is then input into a redundant prediction refinement submodule for extraction to generate residual information used to compensate for information loss.

5. The video steganography method based on hybrid scale reversible attention according to claim 4, characterized in that, The process of inputting the redundant feature representation into the redundancy prediction refinement submodule for extraction is specifically as follows: A feedforward network with dual-branch deep convolutions is used to perform multi-scale filtering on redundant feature representations of the input; information-rich channels are adaptively enhanced and redundant channels are suppressed through a channel attention mechanism; and features are progressively optimized and purified by stacking multiple residual blocks to eliminate irrelevant redundant information while strengthening effective features related to video motion patterns and texture structures.

6. The video steganography method based on hybrid scale reversible attention according to claim 1, characterized in that, The pre-trained end-to-end fully reversible video steganalysis neural network is specifically as follows: End-to-end joint optimization is performed using a multi-objective loss function, which is a weighted sum of a loss module and an enhancement loss module. The loss module includes forward fitting loss, inverse reconstruction loss, and secret recovery loss; the enhancement loss module includes sparse loss, time-aware loss, and time-series gradient smoothing loss.

7. The video steganography method based on hybrid scale reversible attention according to claim 6, characterized in that, Forward fitting loss is used to enforce visual consistency between the stegana video and the cover video, inverse reconstruction loss is used to enforce consistency between the recovered cover video and the original cover video, and secret recovery loss is used to enforce matching between the recovered secret video and the original secret video. Sparse loss is used to constrain the magnitude and distribution of modifications to the original video features during the embedding process, temporal-aware loss is used to utilize the features of the pre-trained spatiotemporal network to maintain inter-frame consistency, and temporal gradient smoothing loss is used to constrain the smooth changes in differences between adjacent frames.

8. A video steganography system based on hybrid scale reversible attention, characterized in that, include: The preprocessing module is used to acquire the cover video frame and the secret video frame to be hidden, and to perform discrete wavelet transform on the two types of video frames respectively to map them to the frequency domain, so as to obtain the cover frequency domain features and the secret frequency domain features. The steganography processing module is used to input the cover frequency domain features and secret frequency domain features into a pre-trained end-to-end fully reversible video steganography neural network, and complete multi-level reversible transformation through multiple cascaded multi-scale hybrid reversible blocks. The transformation result is then input into a dual attention redundancy refinement module to generate residual information. The generation and restoration module is used to reconstruct the transformation result back to the spatial domain through inverse discrete wavelet transform to generate a steganalytic video frame; based on the steganalytic video frame and residual information, the original video features are separated and restored, and then the original video frame at the pixel level is restored through inverse discrete wavelet transform, thus realizing video steganalysis based on hybrid scale reversible attention.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the video steganography method based on hybrid scale reversible attention as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the video steganography method based on hybrid scale reversible attention as described in any one of claims 1 to 5.

Images