U-shaped efficient multi-mode steganography method and system and medium
By using an efficient U-shaped steganography network, combined with parameter sharing and wavelet transform, the problem of performance and complexity imbalance in multimodal steganography is solved, and efficient multimodal information transmission and recovery are achieved.
Patent Information
- Application Number
- CN202511112263.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-11-21
AI Technical Summary
Existing image and speech steganography methods struggle to balance performance, model size, and computational complexity in multimodal information transmission, and they are also ineffective at feature learning at different scales.
An efficient U-shaped steganography network is adopted. Through parameter-sharing forward hiding and backward recovery modules, combined with affine coupling and scaling layers, wavelet transform is used to convert information to the frequency domain for steganography and recovery. Long-range dependencies are established through skip connections, and the interaction function is optimized to save model parameters and computational complexity.
It achieves efficient multimodal steganography with fewer parameters and less computational complexity, improves information fusion and deentanglement capabilities, saves model parameters and computational complexity, and improves steganography and recovery quality.
Smart Images
Figure CN120997024A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a U-shaped high-efficiency multimodal steganography method, system, and medium, belonging to the field of information security technology. Background Technology
[0002] Image steganography enables the covert transmission of information between senders and receivers without being visually detected. This technology not only significantly enhances information security but also demonstrates important application value in several key fields such as medicine, military, and finance. Early research focused primarily on the covert transmission of binary secret information; however, in recent years, with the rapid development of big data and multimedia technologies, research on steganography has gradually shifted from the covert transmission of smaller-capacity binary secret information to the covert transmission of larger-capacity secret images. Simultaneously, with the development of artificial intelligence and speech technology, research on speech information steganography has become increasingly important. However, due to the differences in image and speech characteristics, there is currently no unified steganography paradigm that can simultaneously achieve multimodal steganography. Therefore, against the backdrop of the rapid development of big data and multimedia technologies, how to construct an efficient and unified steganography paradigm to support the covert transmission of different modalities of information has become a critical problem that urgently needs to be solved.
[0003] Currently, image steganography methods can be broadly categorized into two types. One type is based on U-shaped deep neural networks, which requires training two separate networks for each of the secret hiding and recovery processes. Therefore, this type of method has high parameter costs and cannot effectively utilize the knowledge gained from the hiding process to aid in secret recovery. The other type is based on reversible neural networks, which share network parameters during both hiding and recovery, hiding and recovering secret information through the forward and reverse directions of the network. This type of method can effectively conserve model parameters; however, due to the characteristics of reversible neural networks, it is implemented at a single scale, making feature learning impossible at different scales and resulting in excessively high Flops. Therefore, existing image steganography methods struggle to maintain a satisfactory balance between steganography performance, model size, and Flops.
[0004] Therefore, in order to solve the above-mentioned technical problems, there is an urgent need for a U-shaped, efficient, multimodal steganography method, system, and medium. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a U-shaped high-efficiency multimodal steganography method, system and medium. It proposes a high-efficiency U-shaped steganography network that can achieve high-efficiency multimodal steganography with fewer parameters and FLOPS. In the forward and reverse processes of the network, the carrier and secret information are gradually coupled and decoupled from different network depths and scales.
[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:
[0007] In a first aspect, the present invention provides a U-shaped high-efficiency multimodal steganography method, comprising:
[0008] Obtain the carrier diagram and secret information;
[0009] The carrier image and secret information are input into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the secret image output by the high-efficiency U-shaped steganography network model.
[0010] The efficient U-shaped steganography network model includes a forward hiding module for hiding secret information and a backward recovery module for recovering secret information. The forward hiding module and the backward recovery module share parameters. Each of the forward hiding module and the backward recovery module includes n pairs of scaled reversible efficient U-shaped steganography blocks and one unscaled U-shaped reversible block. Skip connections are used between each pair of efficient U-shaped steganography blocks.
[0011] The efficient U-shaped steganography block includes an affine coupling layer and a scaling transformation layer, and the U-shaped reversible block includes an affine coupling layer.
[0012] In the forward hiding module, the affine coupling layer is a fusion layer; in the backward recovery module, the affine coupling layer is a separation layer; and the scale transformation layer is an upsampling layer or a downsampling layer.
[0013] Furthermore, the method also includes:
[0014] Obtain the encrypted image to be decrypted;
[0015] The encrypted image to be decrypted is input into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the carrier image and secret information output by the high-efficiency U-shaped steganography network model.
[0016] Furthermore, the method also includes:
[0017] Wavelet transform is used to convert carrier images and secret information of different modes from the spatial domain to the frequency domain for steganography and recovery;
[0018] The forward hiding process is represented by the following equation:
[0019] (1);
[0020] in This represents an efficient U-shaped steganographic block within the forward hiding module. To discard information, For the steganographic map, 2n+1 is the number of efficient U-shaped steganographic blocks. and These represent wavelet transform and inverse transform, respectively.
[0021] The backward recovery process is represented by the following formula:
[0022] (2);
[0023] in This represents an efficient U-shaped steganography block in the backward recovery module. Auxiliary variables for initializing discarded information.
[0024] Furthermore, in the forward hiding module, when the first The input of an efficient U-shaped steganographic block is , Indicates the height of the carrier image. Indicates the width of the carrier image, the first... Output of a high-efficiency U-shaped steganographic block It can be expressed by the following formula:
[0025] (3);
[0026] in, Represents the downsampling layer. This indicates an upsampling layer. The downsampling layer uses max pooling, and the upsampling layer uses bilinear interpolation. Additionally, a 1×1 convolutional layer is used for downscaling and upscaling of channel dimension features. This represents the affine coupling layer, which acts as the fusion layer during the forward hiding process. This indicates a cascading operation along the channel dimension;
[0027] In the In an efficient U-shaped steganographic block, the input of the affine coupling layer It can be expressed by the following formula:
[0028] (4);
[0029] Corresponding to the input carrier and secret information, the affine coupling layer will Decomposed into and And it is fused through weighted addition and multiplication operations. and to obtain and ;
[0030] No. The outputs of each affine coupling layer are concatenated along the channel dimension. and The resulting splicing has the same characteristics as Same size;
[0031] and Obtained from the following formula:
[0032] (5);
[0033] (6);
[0034] in It represents the Hadamah accumulation. Represents an exponential function. This represents the sigmoid function. , , and The preprocessed interaction functions are shared in the forward hiding and backward recovery processes. These interaction functions have the same structure. At different U-shaped network depths, the carrier information and secret information are gradually coupled through segmentation and weighted fusion operations to generate a secret image.
[0035] Furthermore, in the backward recovery process, the first The input of an efficient U-shaped steganographic block is , No. Output of a high-efficiency U-shaped steganographic block It can be expressed by the following formula:
[0036] (7);
[0037] in This indicates the reverse flow operation of the affine coupling layer, which is a decoupled layer during the backward recovery process.
[0038] In the In a highly efficient U-shaped steganography block, the affine coupling layer The input is represented by the following formula:
[0039] (8);
[0040] During the backward recovery process, the affine coupling layer will Decomposed into and The input is processed by weighted subtraction and division operations. Decoupling, generating output ;
[0041] and The process of obtaining is represented by the following formula:
[0042] (9);
[0043] (10).
[0044] Furthermore, the method for processing the interaction function includes:
[0045] Obtain the channel dimension weights of the input to the interaction function;
[0046] Based on the channel dimension weights, the input of the interaction function is split into a first-priority important part and a second-priority important part;
[0047] Different feature learning methods are applied to the most important and second most important parts to obtain the first learning result and the second learning result.
[0048] The first and second learning results are concatenated along the channel dimension, and the channel order is restored to obtain the final output.
[0049] The priority important part adopts 3 Feature learning is performed using a 3x3 convolutional layer, and the second most important part employs downsampling and 1x3 convolutional layer. 1. Feature learning is achieved through convolution and upsampling;
[0050] Methods for obtaining the channel dimension weights of the interaction function input include:
[0051] The input tensor is pooled to obtain the pooling result;
[0052] Using 1 1. Convolution and activation functions are used to learn the pooling results to obtain the channel dimension weights of the interaction function input.
[0053] Furthermore, the training method for the efficient U-shaped steganography network model includes:
[0054] The sender obtains the carrier map and secret information, and then uses wavelet transform to transform the original carrier map and original secret information into the frequency domain before inputting them into the forward hiding module to obtain the frequency domain carrier map generated by the forward hiding module.
[0055] The generated frequency domain carrier map is transformed to the spatial domain using inverse wavelet transform to obtain the spatial domain carrier map, which is then transmitted from the sender to the receiver.
[0056] The receiver receives the spatial domain secret map, transforms it to the frequency domain using wavelet transform, and then inputs it into the backward recovery module to obtain the frequency domain recovered secret information generated by the backward recovery module.
[0057] The generated frequency domain secret information is transformed into the spatial domain by using inverse wavelet transform to obtain the spatial domain secret information.
[0058] The efficient U-shaped steganography network model is trained based on the spatial domain secret map, the original carrier map, the original secret information, the spatial domain recovered secret information, and the preset loss function to obtain the trained efficient U-shaped steganography network model.
[0059] The loss function is defined as follows:
[0060] (11);
[0061] in, For the total loss, To conceal the losses, To recover the losses, The norm is 2, and M is the number of training samples. To restore the lost weights.
[0062] Secondly, the present invention provides a U-shaped high-efficiency multimodal steganography system, comprising:
[0063] Acquisition module: used to acquire carrier diagrams and secret information;
[0064] Steganography module: used to input the carrier image and secret information into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the steganography image output by the high-efficiency U-shaped steganography network model;
[0065] The efficient U-shaped steganography network model includes a forward hiding module for hiding secret information and a backward recovery module for recovering secret information. The forward hiding module and the backward recovery module share parameters. Each of the forward hiding module and the backward recovery module includes n pairs of scaled reversible efficient U-shaped steganography blocks and one unscaled U-shaped reversible block. Skip connections are used between each pair of efficient U-shaped steganography blocks.
[0066] The efficient U-shaped steganography block includes an affine coupling layer and a scaling transformation layer, and the U-shaped reversible block includes an affine coupling layer.
[0067] In the forward hiding module, the affine coupling layer is a fusion layer; in the backward recovery module, the affine coupling layer is a separation layer; and the scale transformation layer is an upsampling layer or a downsampling layer.
[0068] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.
[0069] Fourthly, the present invention provides a computer device, comprising:
[0070] Memory, used to store computer programs / instructions;
[0071] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.
[0072] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0073] The U-shaped high-efficiency multimodal steganography method provided by this invention proposes an efficient U-shaped steganography network that can achieve efficient multimodal steganography with fewer parameters and FLOPS cost. In the forward and reverse processes of the network, the carrier and secret information are gradually coupled and decoupled from different network depths and scales. In addition, the long-range dependencies generated by skip connections can further improve the steganography and recovery quality.
[0074] The U-shaped high-efficiency multimodal steganography method provided by this invention uses wavelet transform to transform information to the frequency domain to ensure successful steganography of different modal information;
[0075] The U-shaped high-efficiency multimodal steganography method provided by this invention can significantly improve information fusion and deentanglement capabilities, achieve steganography with a shallower multi-scale network, and share parameters during the hiding and recovery processes, thus greatly saving model parameters and computational complexity (Flops).
[0076] The U-shaped efficient multimodal steganography method provided by this invention proposes an efficient interaction function that sorts the importance of input features and performs feature learning on the channel dimension only for the less important features, thereby saving model parameters and computational complexity to a great extent while ensuring the steganography results. Attached Figure Description
[0077] Figure 1 This is a flowchart of a U-shaped high-efficiency multimodal steganography method provided in an embodiment of the present invention;
[0078] Figure 2 This is a system block diagram of the U-shaped high-efficiency multimodal steganography method provided in the embodiments of the present invention;
[0079] Figure 3 It is an efficient interactive function structure diagram;
[0080] Figure 4 This is a schematic diagram of an efficient interactive function convolutional layer. Detailed Implementation
[0081] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.
[0082] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0083] Example 1:
[0084] Figure 1 This is a flowchart of a U-shaped high-efficiency multimodal steganography method according to Embodiment 1 of the present invention. The U-shaped high-efficiency multimodal steganography method provided in this embodiment can be applied to a terminal and can be executed by a U-shaped high-efficiency multimodal steganography system. This system can be implemented by software and / or hardware and can be integrated into the terminal, such as any smartphone, tablet computer, or computer device with communication capabilities. See also... Figure 1 The method implemented in this way specifically includes the following steps:
[0085] Obtain the carrier diagram and secret information;
[0086] The carrier image and secret information are input into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the secret image output by the high-efficiency U-shaped steganography network model.
[0087] The efficient U-shaped steganography network model includes a forward hiding module for hiding secret information and a backward recovery module for recovering secret information. The forward hiding module and the backward recovery module share parameters, which effectively saves the number of parameters. Each of the forward hiding module and the backward recovery module includes n pairs of scaled reversible efficient U-shaped steganography blocks and one unscaled U-shaped reversible block. Skip connections are used between each pair of efficient U-shaped steganography blocks.
[0088] The efficient U-shaped steganography block includes an affine coupling layer and a scaling transformation layer, and the U-shaped reversible block includes an affine coupling layer.
[0089] In the forward hiding module, the affine coupling layer is a fusion layer; in the backward recovery module, the affine coupling layer is a separation layer; and the scale transformation layer is an upsampling layer or a downsampling layer.
[0090] The processing method provided in this embodiment involves the following steps in its application:
[0091] The method further includes:
[0092] Obtain the encrypted image to be decrypted;
[0093] The encrypted image to be decrypted is input into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the carrier image and secret information output by the high-efficiency U-shaped steganography network model.
[0094] The method further includes:
[0095] Wavelet transform is used to convert carrier images and secret information of different modes from the spatial domain to the frequency domain for steganography and recovery;
[0096] The forward hiding process is represented by the following equation:
[0097] (1);
[0098] in This represents an efficient U-shaped steganographic block within the forward hiding module. To discard information, For the steganographic map, 2n+1 is the number of efficient U-shaped steganographic blocks. and These represent wavelet transform and inverse transform, respectively.
[0099] The backward recovery process is represented by the following formula:
[0100] (2);
[0101] in This represents an efficient U-shaped steganography block in the backward recovery module. Auxiliary variables for initializing discarded information.
[0102] In the forward hiding module, when the first The input of an efficient U-shaped steganographic block is , Indicates the height of the carrier image. The width of the carrier graph and the output of the i-th efficient U-shaped steganographic block. It can be expressed by the following formula:
[0103] (3);
[0104] in, Represents the downsampling layer. This indicates an upsampling layer. The downsampling layer uses max pooling, and the upsampling layer uses bilinear interpolation. Additionally, a 1×1 convolutional layer is used for downscaling and upscaling of channel dimension features. This represents the affine coupling layer, which acts as the fusion layer during the forward hiding process. This indicates a cascading operation along the channel dimension;
[0105] In the In an efficient U-shaped steganographic block, the input of the affine coupling layer It can be expressed by the following formula:
[0106] (4);
[0107] Corresponding to the input carrier and secret information, the affine coupling layer will Decomposed into and And it is fused through weighted addition and multiplication operations. and to obtain and ;
[0108] No. The outputs of each affine coupling layer are concatenated along the channel dimension. and The resulting splicing has the same characteristics as Same size;
[0109] and Obtained from the following formula:
[0110] (5);
[0111] (6);
[0112] in It represents the Hadamah accumulation. Represents an exponential function. This represents the sigmoid function. , , and The preprocessed interaction functions are shared in the forward hiding and backward recovery processes. These interaction functions have the same structure. At different U-shaped network depths, the carrier information and secret information are gradually coupled through segmentation and weighted fusion operations to generate a secret image.
[0113] During the backward recovery process, the first The input of an efficient U-shaped steganographic block is , No. Output of a high-efficiency U-shaped steganographic block It can be expressed by the following formula:
[0114] (7);
[0115] in This indicates the reverse flow operation of the affine coupling layer, which is a decoupled layer during the backward recovery process.
[0116] In the In a highly efficient U-shaped steganography block, the affine coupling layer The input is represented by the following formula:
[0117] (8);
[0118] During the backward recovery process, the affine coupling layer will Decomposed into and The input is processed by weighted subtraction and division operations. Decoupling, generating output ;
[0119] and The process of obtaining is represented by the following formula:
[0120] (9);
[0121] (10).
[0122] Furthermore, the method for processing the interaction function includes:
[0123] Obtain the channel dimension weights of the input to the interaction function;
[0124] Based on the channel dimension weights, the input of the interaction function is split into a first-priority important part and a second-priority important part;
[0125] Different feature learning methods are applied to the most important and second most important parts to obtain the first learning result and the second learning result.
[0126] The first and second learning results are concatenated along the channel dimension, and the channel order is restored to obtain the final output.
[0127] The priority important part adopts 3 Feature learning is performed using a 3x3 convolutional layer, and the second most important part employs downsampling and 1x3 convolutional layer. 1. Feature learning is achieved through convolution and upsampling;
[0128] Methods for obtaining the channel dimension weights of the interaction function input include:
[0129] The input tensor is pooled to obtain the pooling result;
[0130] Using 1 1. Convolution and activation functions are used to learn the pooling results to obtain the channel dimension weights of the interaction function input.
[0131] Furthermore, the training method for the efficient U-shaped steganography network model includes:
[0132] The sender obtains the carrier map and secret information, and then uses wavelet transform to transform the original carrier map and original secret information into the frequency domain before inputting them into the forward hiding module to obtain the frequency domain carrier map generated by the forward hiding module.
[0133] The generated frequency domain carrier map is transformed to the spatial domain using inverse wavelet transform to obtain the spatial domain carrier map, which is then transmitted from the sender to the receiver.
[0134] The receiver receives the spatial domain secret map, transforms it to the frequency domain using wavelet transform, and then inputs it into the backward recovery module to obtain the frequency domain recovered secret information generated by the backward recovery module.
[0135] The generated frequency domain secret information is transformed into the spatial domain by using inverse wavelet transform to obtain the spatial domain secret information.
[0136] The efficient U-shaped steganography network model is trained based on the spatial domain secret map, the original carrier map, the original secret information, the spatial domain recovered secret information, and the preset loss function to obtain the trained efficient U-shaped steganography network model.
[0137] The loss function is defined as follows:
[0138] (11);
[0139] in, For the total loss, To conceal the losses, To recover the losses, The norm is 2, and M is the number of training samples. To restore the lost weights.
[0140] The following description, in conjunction with a preferred embodiment, illustrates the content involved in the above embodiments.
[0141] The efficient U-shaped steganography network proposed in this invention consists of two processes: a forward hiding process and a backward recovery process. The overall framework of the proposed efficient U-shaped steganography network is as follows: Figure 2 As shown. Specifically, the forward hiding process and the backward recovery process are as follows:
[0142] 1.1 Forward Hiding Process
[0143] The input to the forward hiding process is the carrier graph. and secret information Cascade .in, This indicates the channel dimension of the carrier diagram. The channel dimension representing secret information. Indicates the height of the carrier image. The width of the carrier image is represented. In the forward pass of the efficient U-shaped steganography network, the carrier and secret information gradually merge and couple. A steganalysis image is generated by combining scaled efficient U-shaped steganography blocks and one unscaled U-shaped reversible block with n (where n represents the number of scaled efficient U-shaped steganography block pairs). Simultaneously, to establish long-range dependencies in the proposed efficient U-shaped steganography network, skip connections are used between each pair of U-shaped steganography blocks. The forward hiding process can be described as follows:
[0144] (1)
[0145] in This indicates a positive, efficient U-shaped steganographic block. To discard information and ensure the security of steganography, only the ciphertext is carried. It can be transmitted to the receiver. 2n+1 is the number of efficient U-shaped steganographic blocks. This invention uses wavelet transform to convert carriers and secret information of different modes into the frequency domain for secret steganography and recovery. and These represent wavelet transform and inverse transform, respectively.
[0146] 1.2 Backward Recovery Process
[0147] The input to the backward recovery process is the received steganalysis image. and initial auxiliary variables Cascade Similar to the forward hiding process, and These are used to convert the carrier and secret information to the frequency domain, and the carrier-secret map to the spatial domain, respectively. The reverse of the forward concealment process is used to recover the carrier information. and recovering secret information Gradually from the reverse of the hiding process Decoupling and reconstruction in the middle.
[0148] The efficient U-shaped steganography block shares some parameters between the forward hiding and backward recovery processes, which effectively saves on the number of parameters. The reverse recovery process can be described as follows:
[0149] (2)
[0150] in This indicates a reverse-efficient U-shaped steganographic block.
[0151] The total loss of this invention consists of hidden losses. and recovery of losses Composition, defined as follows
[0152] (3)
[0153] in It is a 2-norm. M is the number of training samples. To restore the lost weights.
[0154] 2 High-efficiency U-shaped steganography blocks
[0155] The efficient U-shaped steganography block consists of an affine coupling layer and a scaling layer. The affine coupling layer is either a fusion layer or a separation layer, and the scaling layer is either an upsampling layer or a downsampling layer. The structure of the efficient U-shaped steganography block is as follows: Figure 2 The purple trapezoidal block is shown in the image.
[0156] 2.1 Forward Hiding Process
[0157] In the forward hiding process, suppose the first... The input of an efficient U-shaped steganographic block is Then the i-th efficient U-shaped steganographic block The output can be calculated as follows:
[0158] (4)
[0159] in, Represents the downsampling layer. This represents the upsampling layer. The downsampling layer uses max pooling, and the upsampling layer uses bilinear interpolation. Additionally, 1×1 convolutional layers are used for downscaling and upscaling of channel dimension features. This represents the affine coupling layer, which becomes the fusion layer during the forward hiding process. This indicates a cascading operation along the channel dimension. In the... In an efficient U-shaped steganographic block, the input of the affine coupling layer It can be represented as follows:
[0160] (5)
[0161] Corresponding to the carrier and secret information of the network input, the affine coupling layer will Decomposed into and And it is fused through weighted addition and multiplication operations. and to obtain and . No. The output of the affine coupling layer is concatenated along the channel dimension. and The resulting splicing has the same characteristics as Same size. The formula for the affine coupling layer in the forward hiding stage is as follows:
[0162] (6)
[0163] (7)
[0164] in It represents the Hadamah accumulation. Represents an exponential function. This represents the sigmoid function. , , and The interaction functions are shared in both the forward hiding and backward recovery processes, and these interaction functions have the same structure. At different U-shaped network depths, the carrier information and secret information are gradually coupled through segmentation and weighted fusion operations to generate a secret-carrying image.
[0165] 2.2 Backward Recovery Process
[0166] In the backward recovery process, the direction of information flow is from the (2n+1)th efficient U-shaped steganography block to the first efficient U-shaped steganography block, which is the opposite of the forward hiding phase. Assume the... The input of an efficient U-shaped steganographic block is The output is This process can be formalized as follows:
[0167] (8)
[0168] in This represents the reverse flow operation of the affine coupling layer, which acts as a separation layer during the backward recovery process. In the... In a highly efficient U-shaped steganography block, the affine coupling layer The input can be represented as follows:
[0169] (9)
[0170] Corresponding to the weighted addition and weighted multiplication operations of the affine coupling layer during the forward hiding process, the input is weighted by subtraction and division operations during the backward recovery process. Decoupling to generate output The affine coupling layer during the backward recovery process can be represented by the following formula:
[0171] (10)
[0172] (11)
[0173] In each efficient U-shaped steganographic block during the recovery process, the secret and carrier information are progressively decoupled and reconstructed from the carrier image through weighted subtraction and weighted division operations. Simultaneously, the shared parameters between the forward hiding and reverse recovery processes effectively reduce the model's parameters and Flops. Furthermore, due to the multi-scale design of the efficient U-shaped steganographic blocks, the model's Flops will be further reduced.
[0174] 3 Interaction Functions
[0175] To further conserve model parameters, this invention optimizes and improves the classic interaction function structure, proposing an efficient interaction function. The efficient interaction function structure is as follows: Figure 3 As shown, each cube represents a combination of a convolutional layer, a normalization layer, and an activation layer. To conserve model parameters, this invention optimizes the convolutional layer, proposing an efficient interaction function convolutional layer, the structure of which is shown below. Figure 4 As shown.
[0176] To further improve steganography efficiency, the proposed efficient interaction function improves the convolutional layer in the classic interaction function, as follows:
[0177] To obtain the weights in the channel dimension First, pooling is used to encode each feature map. The pooling kernel sizes are as follows: When the input is Pooling operations can be represented by the following formula:
[0178] (12)
[0179] Among them, pooling output respectively through pooling kernel The output of the pooling operation. Then, using... Convolutional layer The obtained pooling results are then further processed. This process can be represented by the following formula:
[0180] (13)
[0181] in, Represents the nonlinear ReLU activation function. This is the initial weight.
[0182] Then, adopt Convolutional layer right Further learning yields the final weights. This process can be represented by the following formula:
[0183] (14)
[0184] Then, the input X (which is the input of the radial coupling layer mentioned above) is weighted according to the channel weights. Or the input of the separation layer Sort by channel dimension, and then break down by channel dimension according to importance, dividing into priority and important parts. and second priority part The splitting process is based on the channel importance ratio. Proceed, prioritizing important parts The proportion of the total channel input X is Second priority part The proportion of the total channel input X is For priority and important parts Using the original Convolutional layer Feature learning can be represented by the following formula:
[0185] (15)
[0186] For the second priority part Only its channel characteristics are learned, and it is downsampled. Convolutional layers perform feature learning, followed by downsampling for scale recovery. This process can be represented by the following formula:
[0187] (16)
[0188] Finally, the final output is obtained by concatenating the results of the two learning parts along the channel dimension and restoring the channel order. This process can be represented by the following formula:
[0189] (17)
[0190] in, This indicates a channel order restoration operation. This indicates a channel dimension splicing operation.
[0191] The specific training process of the high-efficiency U-shaped steganography network provided above is as follows:
[0192] Initialization parameters: Number of secret images: And the number of carrier images: ;
[0193] Number of efficient U-shaped steganographic blocks: 2n+1;
[0194] Network parameters:
[0195] Recovery phase weights in the loss function: ;
[0196] Channel importance ratio ;
[0197] Repeat the following steps (within one epoch) until the network converges:
[0198] Input for the steganography stage: original carrier graph and original secret information ;
[0199] Wavelet transform is used to transform the original carrier image and the original secret information to the frequency domain;
[0200] Generate the confidentiality map using the following formula. :
[0201] ;
[0202] The generated density map is transformed to the spatial domain using inverse wavelet transform to obtain the spatial domain density map.
[0203] The sender will transmit the spatial domain-based encrypted map;
[0204] Receiver receives a spatial domain data map;
[0205] The received carrier map is transformed to the frequency domain using wavelet transform;
[0206] Generate recovery secret information according to the following formula. :
[0207] ;
[0208] The secret information is transformed into the spatial domain using inverse wavelet transform;
[0209] Calculate the loss function and update the parameters according to the following formula. : .
[0210] Example 2: This example provides a U-shaped high-efficiency multimodal steganography system, including:
[0211] Acquisition module: used to acquire carrier diagrams and secret information;
[0212] Steganography module: used to input the carrier image and secret information into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the steganography image output by the high-efficiency U-shaped steganography network model;
[0213] The efficient U-shaped steganography network model includes a forward hiding module for hiding secret information and a backward recovery module for recovering secret information. The forward hiding module and the backward recovery module share parameters. Each of the forward hiding module and the backward recovery module includes n pairs of scaled reversible efficient U-shaped steganography blocks and one unscaled U-shaped reversible block. Skip connections are used between each pair of efficient U-shaped steganography blocks.
[0214] The efficient U-shaped steganography block includes an affine coupling layer and a scaling transformation layer, and the U-shaped reversible block includes an affine coupling layer.
[0215] In the forward hiding module, the affine coupling layer is a fusion layer; in the backward recovery module, the affine coupling layer is a separation layer; and the scale transformation layer is an upsampling layer or a downsampling layer.
[0216] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.
[0217] Example 3: This example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in Example 1.
[0218] Example 4: This example provides a computer device, including:
[0219] Memory, used to store computer programs / instructions;
[0220] A processor for executing the computer program / instructions to implement the steps of any of the methods described in Embodiment 1.
[0221] Example 5: This example provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any one of Examples 1.
[0222] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
[0223] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure 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.
[0224] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. 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, create a machine 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.
[0225] 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.
[0226] 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.
[0227] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.
Claims
1. A U-shaped, efficient, multimodal steganography method, characterized in that, include: Obtain the carrier diagram and secret information; The carrier image and secret information are input into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the secret image output by the high-efficiency U-shaped steganography network model. The efficient U-shaped steganography network model includes a forward hiding module for hiding secret information and a backward recovery module for recovering secret information. The forward hiding module and the backward recovery module share parameters. Each of the forward hiding module and the backward recovery module includes n pairs of scaled reversible efficient U-shaped steganography blocks and one unscaled U-shaped reversible block. Skip connections are used between each pair of efficient U-shaped steganography blocks. The efficient U-shaped steganography block includes an affine coupling layer and a scaling transformation layer, and the U-shaped reversible block includes an affine coupling layer; In the forward hiding module, the affine coupling layer is a fusion layer; in the backward recovery module, the affine coupling layer is a separation layer; and the scale transformation layer is an upsampling layer or a downsampling layer.
2. The U-shaped high-efficiency multimodal steganography method according to claim 1, characterized in that, The method further includes: Obtain the encrypted image to be decrypted; The encrypted image to be decrypted is input into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the carrier image and secret information output by the high-efficiency U-shaped steganography network model.
3. The U-shaped high-efficiency multimodal steganography method according to claim 2, characterized in that, The method further includes: Wavelet transform is used to convert carrier images and secret information of different modes from the spatial domain to the frequency domain for steganography and recovery; The forward hiding process is represented by the following equation: (1)； in This represents an efficient U-shaped steganographic block within the forward hiding module. To discard information, For the steganographic map, 2n+1 is the number of efficient U-shaped steganographic blocks. and These represent wavelet transform and inverse transform, respectively. The backward recovery process is represented by the following formula: (2)； in This represents an efficient U-shaped steganography block in the backward recovery module. Auxiliary variables for initializing discarded information.
4. The U-shaped high-efficiency multimodal steganography method according to claim 3, characterized in that, In the forward hiding module, when the first The input of an efficient U-shaped steganographic block is , Indicates the height of the carrier image. Indicates the width of the carrier image, the first... Output of a high-efficiency U-shaped steganographic block It can be expressed by the following formula: (3)； in, Represents the downsampling layer. This indicates an upsampling layer. The downsampling layer uses max pooling, and the upsampling layer uses bilinear interpolation. Additionally, a 1×1 convolutional layer is used for downscaling and upscaling of channel dimension features. This represents the affine coupling layer, which acts as the fusion layer during the forward hiding process. This indicates a cascading operation along the channel dimension; In the In an efficient U-shaped steganographic block, the input of the affine coupling layer It can be expressed by the following formula: (4)； Corresponding to the input carrier and secret information, the affine coupling layer will Decomposed into and And it is fused through weighted addition and multiplication operations. and to obtain and ; No. The outputs of each affine coupling layer are concatenated along the channel dimension. and The resulting splicing has the same characteristics as Same size; and Obtained from the following formula: (5)； (6)； in It represents the Hadamah accumulation. Represents an exponential function. This represents the sigmoid function. , , and The preprocessed interaction functions are shared in the forward hiding and backward recovery processes. These interaction functions have the same structure. At different U-shaped network depths, the carrier information and secret information are gradually coupled through segmentation and weighted fusion operations to generate a secret image.
5. The U-shaped high-efficiency multimodal steganography method according to claim 4, characterized in that, During the backward recovery process, the first The input of an efficient U-shaped steganographic block is , No. Output of a high-efficiency U-shaped steganographic block It can be expressed by the following formula: (7)； in This indicates the reverse flow operation of the affine coupling layer, which is a decoupled layer during the backward recovery process. In the In a highly efficient U-shaped steganography block, the affine coupling layer The input is represented by the following formula: (8)； During the backward recovery process, the affine coupling layer will Decomposed into and The input is processed by weighted subtraction and division operations. Decoupling, generating output ; and The process of obtaining is represented by the following formula: (9)； (10)。 6. The U-shaped high-efficiency multimodal steganography method according to claim 5, characterized in that, The method for processing the interaction function includes: Obtain the channel dimension weights of the input to the interaction function; Based on the channel dimension weights, the input of the interaction function is split into a first-priority important part and a second-priority important part; Different feature learning methods are applied to the most important and second most important parts to obtain the first learning result and the second learning result. The first and second learning results are concatenated along the channel dimension, and the channel order is restored to obtain the final output. The priority important part adopts 3 Feature learning is performed using a 3x3 convolutional layer, and the second most important part employs downsampling and 1x3 convolutional layer.
1. Feature learning is achieved through convolution and upsampling; Methods for obtaining the channel dimension weights of the interaction function input include: The input tensor is pooled to obtain the pooling result; Using 1 1. Convolution and activation functions are used to learn the pooling results to obtain the channel dimension weights of the interaction function input.
7. The U-shaped high-efficiency multimodal steganography method according to claim 6, characterized in that, The training method for the efficient U-shaped steganography network model includes: The sender obtains the original carrier map and the original secret information, and then uses wavelet transform to transform the original carrier map and the original secret information into the frequency domain before inputting them into the forward hiding module to obtain the frequency domain carrier map generated by the forward hiding module. The generated frequency domain carrier map is transformed to the spatial domain using inverse wavelet transform to obtain the spatial domain carrier map, which is then transmitted from the sender to the receiver. The receiver receives the spatial domain secret map, transforms it to the frequency domain using wavelet transform, and then inputs it into the backward recovery module to obtain the frequency domain recovered secret information generated by the backward recovery module. The generated frequency domain secret information is transformed into the spatial domain by using inverse wavelet transform to obtain the spatial domain secret information. The efficient U-shaped steganography network model is trained based on the spatial domain secret map, the original carrier map, the original secret information, the spatial domain recovered secret information, and the preset loss function to obtain the trained efficient U-shaped steganography network model. The loss function is defined as follows: (11)； in, For the total loss, To conceal the losses, To recover the losses, The norm is 2, and M is the number of training samples. To restore the lost weights.
8. A U-shaped high-efficiency multimodal steganography system, characterized in that, include: Acquisition module: used to acquire carrier diagrams and secret information; Steganography module: used to input the carrier image and secret information into a pre-designed and trained high-efficiency U-shaped steganography network model to obtain the steganography image output by the high-efficiency U-shaped steganography network model; The efficient U-shaped steganography network model includes a forward hiding module for hiding secret information and a backward recovery module for recovering secret information. The forward hiding module and the backward recovery module share parameters. Each of the forward hiding module and the backward recovery module includes n pairs of scaled reversible efficient U-shaped steganography blocks and one unscaled U-shaped reversible block. Skip connections are used between each pair of efficient U-shaped steganography blocks. The efficient U-shaped steganography block includes an affine coupling layer and a scaling transformation layer, and the U-shaped reversible block includes an affine coupling layer; In the forward hiding module, the affine coupling layer is a fusion layer; in the backward recovery module, the affine coupling layer is a separation layer; and the scale transformation layer is an upsampling layer or a downsampling layer.
9. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method according to any one of claims 1-7.
10. A computer device, comprising: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the method according to any one of claims 1-7.