A method of cognitive covert communication

Abstract

A cognitively covert communication method belongs to the field of artificial intelligence wireless communication. The implementation method of this invention is as follows: constructing a basic dataset and dataset 1; using optical flow, obtaining optical flow output images through an optical flow field; training an image interpolation neural network to obtain a trained deep learning method; inputting the data from step one into the data from step three to obtain the deep learning output image; putting the image interpolation data into dataset 1 to obtain dataset 2, which is used as training data for a semantic communication network; constructing a semantic communication network; using dataset 2 as training data for the semantic communication network; distributing a covert communication classification network and a semantic communication network, and implementing cognitive covert communication based on the covert communication classification network and the semantic communication network. In the process of training the "semantic consensus" in semantic communication, this invention trains two semantic decoders, each containing different "semantic consensuses," thereby improving the security of covert communication.

Description

Technical Field

[0001] This invention relates to a cognitive covert communication method based on semantic communication, belonging to the field of artificial intelligence wireless communication. Technical Background

[0002] Currently, countries worldwide are paying close attention to the development of cyberspace security. Meanwhile, the synergistic effect of the internet, the multiplier effect of data elements, and the exponential effect of artificial intelligence are important driving forces. Following the trend of technological development, new technologies bring new operational dimensions and transmission methods. The importance of secure information transmission is rising daily, and information concealment is key to achieving secure transmission of critical information. Information concealment refers to a secure information transmission method that hides critical information within normal carriers in cyberspace or electromagnetic space, preventing it from being discovered and intercepted by non-cooperative parties.

[0003] Future information concealment requires comprehensive protection across multiple levels, including signals, devices, and behaviors, necessitating the multi-dimensional development of information concealment technologies. While semantic communication is still in its early stages of research, significant progress has been made in areas such as text-modal end-to-end semantic communication, image-modal end-to-end semantic communication, video-modal end-to-end semantic communication, multimodal end-to-end semantic communication, and the distribution of semantic communication resources. Semantic communication is poised to become a crucial form of communication. Current information concealment aims to ensure that transmitted content and communication behaviors remain undetected, focusing on both the source and channel levels. Source concealment typically employs methods such as hiding, disguising, and deception to conceal information. Channel concealment utilizes methods like domain expansion, dimensional expansion, and integration to conceal the transmission channel. Covert communication through semantic communication channels aligns with technological trends, enabling information concealment that is unpredictable and difficult for adversaries to detect, achieving efficient and undetectable data embedding at the cognitive level. Summary of the Invention

[0004] The purpose of this invention is to achieve information concealment in semantic communication, and to provide a cognitive covert communication method based on semantic communication. This method utilizes semantic-level information for communication, thereby improving communication efficiency. No additional hidden information is added during normal communication; instead, semantic decoders with different semantic consensuses are trained during the training of semantic consensus, making the covert communication behavior difficult to detect.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention discloses a cognitively concealed communication method, characterized by comprising the following steps:

[0007] Step 1: Construct dataset 1.

[0008] Step 2: Use the optical flow method to obtain the optical flow output image through the optical flow field.

[0009] Step 3: Train the image interpolation neural network to obtain a trained deep learning method.

[0010] Step 4: Input the data from Step 1 into the data from Step 3 to obtain the deep learning output image.

[0011] Step 5: Put the image interpolation data into dataset 1 from step 1 to obtain dataset 2, which will be used as training data for the semantic communication network.

[0012] Step 6: Construct a semantic communication network.

[0013] Step 6.1: Encoder construction.

[0014] Step 6.2: Channel construction.

[0015] Step 6.3: Decoder construction.

[0016] Step 7: Use dataset 2 from step 5 as training data for the semantic communication network to train the semantic communication network in step 6.

[0017] Step 7.1: Parameter initialization.

[0018] Step 7.2: Define the loss function.

[0019] Step 7.3: Backpropagation step.

[0020] Step 7.4: Parameter update and optimizer steps.

[0021] Step 7.5: Iterative training.

[0022] Step 8: Distribute the covert communication classification network and semantic communication network, and realize cognitive covert communication based on the covert communication classification network and semantic communication network.

[0023] Furthermore, the specific implementation method of step one is as follows:

[0024] The base dataset consists of a temporally ordered image data stream, with each image annotated with a sequential label. Furthermore, Gaussian smoothing is used to construct the training and test sets. The labels contain only the sequence number of the current image.

[0025] Based on the basic dataset, hidden information is added at the data preparation level by modifying the labels of pre-specified images, thus constructing dataset 1 containing hidden information.

[0026] Furthermore, the specific implementation method for step two is as follows:

[0027] Based on dataset 1 obtained in step one, the optical flow method is used to obtain the optical flow output image through the optical flow field.

[0028] Furthermore, the specific implementation method for step three is as follows:

[0029] Input dataset 1 contains two adjacent RGB images I1 and I2, where I1, I2 ∈ RH×W×3, H is the image height, W is the image width, and 3 represents the number of RGB channels. Perform a PixelUnshuffle downsampling operation on each frame to convert the spatial dimension to the channel dimension, obtaining the downsampled features. and Its dimensions are (H / s)×(W / s)×(3·s) 2 ), where s is the scaling factor. I1 and I2 are concatenated along the channel dimension to generate the merged feature.

[0030] Features are merged using a num×num convolutional layer. The number of channels is compressed to 3. 2 ,get initial features F in Input is sent to a residual group (ResGroup), which contains multiple residual channel attention blocks (RCA Blocks). Each RCA Block performs the following operations: for F in Perform two 3×3 convolutions and ReLU activations to generate intermediate features. Then, apply channel attention weighting to these intermediate features to obtain weighted features. Finally, combine the weighted features with F... in The final feature is output by adding the residuals together.

[0031] The method for obtaining weighted features by performing channel attention weighting is as follows: global average pooling is used to generate channel feature descriptors. Where C represents the number of channels. Channel attention weights Catt(F) are generated through two fully connected layers. c )∈[0,1] C The formula is: att(F c )=σ(W1·ReLU(W0·F c ))att(Fc)

[0032] in Let r be the learnable weight matrix, r be the compression ratio, and σ be the sigmoid function. The intermediate features are then compared with att(F... c Multiply each channel to obtain the weighted feature.

[0033] The final feature count is adjusted by performing a 3×3 convolution. Then, the final feature channel count is converted to 3×3 using a PixelShuffle upsampling operation. 2 And reorganize the spatial dimensions according to s×s blocks to generate intermediate frames.

[0034] Calculate intermediate frames With real intermediate frame I gt Pixel reconstruction loss: in These are actual intermediate frames.

[0035] Calculate the perceptual loss based on the pre-trained network and extract feature differences: Where φ is the feature extraction function of the pre-trained network.

[0036] Jointly optimize the total loss function:

[0037]

[0038] Iterative optimization to The requirements are met, and the training is completed to obtain a well-trained neural network.

[0039] Furthermore, the specific implementation method for step four is as follows:

[0040] Input the dataset 1 from step one into the neural network trained in step three to obtain the output image of the neural network.

[0041] Furthermore, the specific implementation method of step five is as follows:

[0042] The hidden information is encoded according to the label values ​​in dataset 1, resulting in an image stream with encoded hidden information. The optical flow output image obtained in step 2 or the neural network output image obtained in step 4 is inserted into the image stream in step 5. Several images are inserted between every two adjacent images in the image stream. This yields the image interpolation data.

[0043] The image interpolation data is put into dataset 1 to obtain a new dataset, which is used as the training data for the semantic communication network.

[0044] Furthermore, the specific implementation method for step six is ​​as follows:

[0045] Step 6.1: Construct the encoder. The encoder encodes the source data into a format suitable for transmission, while the decoder decodes the received signal back into the source data. The encoder's input layer receives the source data, and the input convolutional layer extracts features from the source data. A Transformer layer captures long-range dependency features from the features extracted by the convolutional layers. The latent space mapping layer maps the long-range dependency features extracted by the Transformer layer to the latent space, obtaining a latent feature vector. This latent feature vector serves as the vector to be transmitted over the channel.

[0046] Step 6.2: Construct a channel model between the encoder and decoder to simulate noise and interference in a real communication environment. The channel model is an additive white Gaussian noise (AWGN) channel. By introducing the channel model during training, the network can adapt to different communication environments. The additive white Gaussian noise (AWGN) channel is...

[0047] y = hx + n

[0048] y represents the received signal, i.e., the signal received by the receiver after transmission through the channel. x represents the latent eigenvector, i.e., the original signal transmitted by the transmitter. h represents the channel gain, indicating the attenuation or amplification effect of the channel on the signal. In practical communication, n represents noise.

[0049] Step 6.3: Construct the encoder. The decoder starts with the received signal and gradually converts it into data suitable for reconstruction. The received signal first passes through the latent space decoding layer to obtain a feature representation suitable for decoding. The deconvolution layer gradually restores the feature representation from the previous step to obtain the deconvolution result. The deconvolution result is then passed through the output layer to generate the reconstructed data.

[0050] The overall encoder-decoder channel model is as follows:

[0051]

[0052] s*: Represents the image reconstructed at the receiving end.

[0053] D θ : Represents the decoder, with parameter θ, used to decode the encoded signal back into an image.

[0054] ω: Represents a function used to process the latent variable z.

[0055] z: represents encoder E φ The extracted latent features have a parameter of φ.

[0056] E φ : Represents the encoder, with parameter φ, used to extract latent features from the original image s. E φ The task is to extract latent features from the original image. s passes through E φ The encoder extracts the source data into latent features through a multi-layer neural network. These features contain the core semantic information of the data. The semantic information is presented as a continuous distribution during the initial extraction process. z passes through channel ω and is received by the receiver, which decodes the encoded signal back into the source data. The decoder then gradually reconstructs the data through a multi-layer neural network. Each layer performs a non-linear transformation on the input to approximate the representation s* of the source data.

[0057] Furthermore, the specific implementation method of step seven is as follows:

[0058] Step 7.1: Before training the semantic communication network, initialize the network parameters. Parameters include the parameters of the encoder, decoder, and channel model. Initialization methods include random initialization or initialization using the parameters of a pre-trained model.

[0059] Step 7.2: Define the network's loss function, using cross-entropy loss (CE) to measure the difference between predicted and true values. Using dataset 2 as training and validation data, for the semantic communication network without hidden information, a source data autoregression approach is used, with CE as the loss function to train the semantic communication network. For the semantic communication network containing hidden information, source data and labels are used, with CE as the loss function to train a classification network to identify hidden information.

[0060] CE is represented as

[0061]

[0062] H(P): Represents information entropy, used to measure the uncertainty of a random variable X. The higher the information entropy, the higher the uncertainty of the random variable.

[0063] P(X=x i ): represents the probability that a random variable X takes the value xi.

[0064] m: The total number of all possible values ​​that the random variable X can take.

[0065] Step 7.3: Forward propagation computation of dataset 2, while simultaneously calculating the gradient of the loss function with respect to the network parameters using the backpropagation algorithm. This process starts from the output layer of the decoder and proceeds layer by layer forward until the input layer of the encoder.

[0066] Step 7.4: Use the Adam optimizer to update the network parameters of the semantic communication network and the hidden communication classification network based on the calculated gradients.

[0067] Step 7.5: Repeat the forward propagation, loss calculation, backpropagation, and parameter update process until the network performance reaches a predetermined threshold or the training reaches a predetermined number of iterations. In each iteration, the network gradually learns better parameters, improving its semantic understanding of the source data, and the training hyperparameters are modified multiple times to improve the training results. After completing multiple iterations of training, a dedicated semantic communication network is obtained.

[0068] Through steps 7.1 to 7.5, the trained semantic communication network can achieve efficient and robust semantic information transmission in complex communication environments, and is suitable for various practical application scenarios.

[0069] Furthermore, the specific implementation method of step eight is as follows:

[0070] Step 8.1: The public distribution of the Semantic Communication Network is carried out through multiple channels.

[0071] Step 8.2: The covert distribution of the classified network is strictly limited to partners, ensuring that only authorized entities can access and use the network.

[0072] Step 8.3: The sender of the covert communication system encodes the pre-transmitted covert information and performs interpolation to obtain the transmission sequence, which is then transmitted on a public channel. After transmission, the receiver of the covert communication system decodes the information to obtain the covert information.

[0073] Beneficial effects:

[0074] 1. This invention discloses a cognitive covert communication method based on semantic communication. It adopts a "understand first, transmit later" semantic communication paradigm, extracting features, compressing and encoding the original signal, and transmitting it under the premise that both communicating parties have "semantic consensus." Compared with traditional communication methods, this method utilizes semantic-level information for communication, improving communication efficiency. During training, two decoders with different "semantic consensus" are formed; one is publicly distributed, and the other is secretly distributed, making covert communication difficult to detect and enhancing communication security.

[0075] 2. This invention discloses a cognitive covert communication method based on semantic communication, employing a differentiated distribution strategy: the semantic communication network distributes publicly, while the covert communication classification network distributes secretly. The public distribution of the semantic communication network expands its application scope, while the secret distribution of the covert communication classification network ensures information security in specific collaborative scenarios, thus balancing system security, privacy, and broad applicability.

[0076] 3. This invention discloses a cognitive covert communication method based on semantic communication, which constructs an additive white Gaussian noise channel model to enable the network to adapt to different communication environments. By introducing the channel model during the training process, compared with communication methods that do not consider channel noise, this method improves the robustness of communication, ensures stable transmission of semantic information in complex environments, and enhances anti-interference capabilities.

[0077] 4. This invention discloses a cognitive covert communication method based on semantic communication. Instead of adding hidden information during normal communication, it trains semantic decoders containing different semantic consensuses during the training of "semantic consensus." By modifying the labels of specific images, covert information is added at the data preparation level without modifying the images themselves. This makes the covert communication behavior difficult to detect, and non-partners cannot distinguish whether covert information is present, further enhancing its covertness.

[0078] 5. This invention discloses a cognitive covert communication method based on semantic communication. Depending on the specific circumstances, it selects either a traditional compensation interpolation method or a deep learning-based image interpolation method to add supplementary images to the base image stream after the covert information is encoded. Compared to a single interpolation method, this method is applicable to different scenarios, reduces the risk of anomaly detection, and improves the reliability of covert communication.

[0079] 6. This invention discloses a cognitive covert communication method based on semantic communication, which utilizes a Transformer network to extract long-range dependency features of source data in the encoder. Because the Transformer network can effectively capture long-range dependency feature information, this method provides a more solid foundation for cognitive covert communication compared to traditional neural networks, improving the accuracy and efficiency of communication. Attached Figure Description

[0080] Figure 1 This is a flowchart of a cognitive covert communication method based on semantic communication disclosed in this invention.

[0081] Figure 2 This is an overall framework diagram of a cognitive covert communication method based on semantic communication disclosed in this invention.

[0082] Figure 3 A schematic diagram of the basic dataset.

[0083] Figure 4 This is a structural diagram of a deep learning frame interpolation model.

[0084] Figure 5 This is a diagram of the semantic communication network structure.

[0085] Figure 6 VIT model structure diagram. Detailed Implementation

[0086] To enable those skilled in the art to more deeply understand the implementation ideas of the present invention, the technical solutions of the present invention will be carefully and clearly described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other implementation cases obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention. The specific steps of the embodiments of the present invention will be described below in conjunction with specific scenarios, and the overall process is as follows: Figure 1 The overall structure diagram shown is as follows: Figure 2 As shown.

[0087] Example:

[0088] This experiment was conducted on a server with the following configuration: Intel(R) NUC 13 Extreme, CPU i9-13900K 2.60GHz, GPU NVIDIA GeForce RTX 3080 32GB, and operating system Windows 11 64-bit.

[0089] like Figure 2 As shown in the figure, this embodiment discloses a cognitive covert communication method based on semantic communication, and the specific implementation steps are as follows:

[0090] Step 1: Construct the dataset.

[0091] The base dataset was constructed using Global Forecast System (GFS) NetCDF wind field data. GFS is a global weather forecasting model from the U.S. National Weather Service (NWS) used to predict global weather changes. NetCDF (Network Common Data Form) is a self-describing, platform-independent data format used for storing and managing scientific data. NetCDF files are commonly used to store meteorological data such as temperature, humidity, and wind speed, which are multidimensional, meaning they vary over time and space. A custom image dataset for the western Pacific region was constructed using GFS NetCDF wind field data. This dataset covers the period from May 2020 to April 2021 and contains 1152 images, each annotated with a sequential label. Each image is normalized to a resolution of 255*255 pixels and has three color channels. The labels on the base dataset only contain the sequence number of the current image. In the end-to-end network training for semantic communication, the dataset is used in an autoregressive manner, meaning the neural network uses reconstruction error as the loss. The training process is described in step three.

[0092] Based on a base dataset, the distribution of data is altered by modifying the dataset's labels, thereby constructing a dataset containing hidden information. Hidden information is added to specific images at the data preparation level by modifying their labels. During actual communication, no hidden information is added to the images, allowing for more effective concealment of covert communication activities. Since the images are not modified, from the perspective of non-partners, it is impossible to distinguish whether hidden information or covert communication activities exist. This section contains 1152 images as base images and also as information-hiding images. The base dataset uses a temporally ordered image data stream, with each image annotated with a sequential label. The label only contains the sequence number of the current image.

[0093] Based on the basic dataset, hidden information is added at the data preparation level by modifying the labels of pre-randomized images, thus constructing dataset 1 containing hidden information.

[0094] Step 2: Based on the dataset 1 obtained in Step 1, use the optical flow method to obtain the optical flow output image through the optical flow field.

[0095] Dataset 1 contains 36 different labels from 0-9 and az. Dividing the basic dataset into 32 subsets, if the label information is directly used as the encoding sequence, non-cooperative parties can observe a disordered image stream, easily leading to anomaly detection risks. Therefore, supplementary images need to be added between the basic images in the image stream after the hidden information is encoded; this is image interpolation, reducing the risk of anomaly detection. This section includes two image interpolation algorithms: traditional compensation interpolation and deep learning-based image interpolation. The interpolation method used depends on the specific situation. Optical flow method: suitable for scenarios with high real-time requirements; deep learning-based method: suitable for scenarios requiring high precision.

[0096] Step 2.1 Image Preprocessing

[0097] The Lucas-Kanad optical flow method is used to describe the motion of each pixel in an image sequence over time. Input: Image I at time t. t Image I at time t+τ t+τ Perform image preprocessing, for I t and I t+τ A suitable kernel size is selected for Gaussian smoothing to reduce the impact of image noise on optical flow estimation and improve the stability of gradient calculation. The Gaussian smoothing formula is: I smooth =I*G, where I represents the original image, G represents the convolution kernel, and * represents the convolution operation. G is defined as:

[0098]

[0099] Where σ represents the standard deviation of the Gaussian kernel, which determines the degree of smoothing.

[0100] Step 2.2 Optical Flow Calculation

[0101] Based on image characteristics and application requirements, the sliding window parameters in the Lucas-Kanade algorithm are carefully selected, covering both fixed and adaptive window modes, window size, and the density and sparsity of optical flow. To calculate the optical flow constraint equation, the Sobel operator is used to compute the spatial gradient of the image in x and y coordinates. With time gradient The Sobel operator is defined as follows:

[0102]

[0103] The temporal gradient can be approximated by the difference between two frames of images:

[0104]

[0105] Among them: I t (x,y) is the pixel intensity of the image at position (x,y) at time t, I t+τ (x,y) is the pixel intensity of the image at position (x,y) at time t+τ.

[0106] In optical flow estimation, temporal gradient and spatial gradient and Together, they form the optical flow constraint equations. Within a selected window region, a set of optical flow constraint equations is constructed to solve for a unified optical flow vector (u,v). The optical flow constraint equations assume that the brightness of each pixel in the image remains constant during motion, i.e.:

[0107]

[0108] in: and These are the spatial gradients of the image in the x and y directions, respectively. u and v are the components of the optical flow in the x and y directions, respectively. This refers to the temporal gradient. Once the optical flow vector (u,v) is obtained, it is precisely assigned to the pixel at the center of the window, thus completing the estimation of local optical flow. Further, by traversing the entire image and repeating the above steps for each pixel where optical flow needs to be calculated, an optical flow field covering the entire image is finally constructed, providing crucial motion information support for subsequent image processing and analysis tasks. Based on dataset 1 obtained in step one, the optical flow method is used to obtain the optical flow output image through the optical flow field.

[0109] Step 3: Based on the dataset 1 obtained in Step 1, train the deep learning method to obtain the trained method.

[0110] Traditional methods process errors step by step, which accumulate at each step. Deep learning methods, on the other hand, optimize the process end-to-end to generate more realistic intermediate frames. By using multiple frames of input to capture long-distance dependencies, end-to-end training allows the model to jointly optimize all steps (from input frames to output frames), reducing intermediate errors and improving the overall frame interpolation quality.

[0111] Input dataset 1 contains two adjacent RGB images I1 and I2, where I1, I2 ∈ RH×W×3, H is the image height, W is the image width, and 3 represents the number of RGB channels. Perform a PixelUnshuffle downsampling operation on each frame to convert the spatial dimension to the channel dimension, obtaining the downsampled features. and Its dimensions are (H / s)×(W / s)×(3·s) 2 ), where s is the scaling factor. I1 and I2 are concatenated along the channel dimension to generate the merged feature.

[0112] Feature compression and residual channel attention processing

[0113] Features are merged using a 3×3 convolutional layer. The number of channels is compressed to 3. 2 To obtain initial features F in The input is fed into a residual group (ResGroup), which contains multiple residual channel attention blocks (RCA Blocks). Each RCA Block performs the following operations: performs two 3×3 convolutions and ReLU activations on the input features to generate intermediate features; and applies channel attention (CA) weighting to the intermediate features, including global average pooling to generate channel descriptors. Where C represents the number of channels. Channel attention weights Catt(F) are generated through two fully connected layers (equivalent to 1×1 convolutions). c )∈[0,1] C The formula is: att(F c )=σ(W1·ReLU(W0·F c ))att(Fc)

[0114] in The learnable weight matrix is ​​defined by r = 16, where r is the compression ratio and σ is the Sigmoid function. The intermediate features are then compared with att(F... c The weighted features are obtained by multiplying each channel. The weighted features are then added to the input features through a residual connection to output the final features.

[0115] Intermediate frame reconstruction and upsampling

[0116] The feature output from the residual group is subjected to a 3×3 convolution to adjust the number of channels. The feature channel number is then converted to 3×3 using a PixelShuffle upsampling operation (scaling factor s = 8). 2 And reorganize the spatial dimensions according to s×s blocks to generate intermediate frames.

[0117] Joint loss function optimization

[0118] Calculate intermediate frames Pixel reconstruction loss compared to the real frame Igt: in These are actual intermediate frames.

[0119] Calculate the perceptual loss based on the pre-trained VGG-19 network and extract the feature differences from the conv5_4 layer: Where φ is the feature extraction function of VGG-19.

[0120] Jointly optimize the total loss function:

[0121] Among them, α=0.9 and β=0.005.

[0122] Iterative optimization to The requirements are met, and the training is completed to obtain a well-trained neural network.

[0123] Step 4: Input the dataset 1 from Step 1 into the neural network trained in Step 3 to obtain the output image of the neural network.

[0124] Step 5: Encode the hidden information according to the label values ​​in dataset 1 to obtain the image stream with encoded hidden information. Insert the optical flow output image obtained in step 2 or the neural network output image obtained in step 4 into the image stream of step 5. Insert several images into every two adjacent images in the image stream. Obtain image interpolation data.

[0125] The image interpolation data is put into dataset 1 to obtain dataset 2, which is used as training data for the semantic communication network.

[0126] Step 6: Construct a semantic communication network

[0127] Step 6.1 Encoder Construction

[0128] Design an encoder. Data from dataset 2 is used as the source data. The encoder encodes the source data into a format suitable for transmission, while the decoder decodes the received signal back into the source data. The encoder's input layer receives the source data, and the input convolutional layer extracts features from the source data. A Transformer layer is used to capture long-range dependency features from the features extracted by the convolutional layers. A latent space mapping layer maps the long-range dependency features extracted by the Transformer layer to the latent space, obtaining a latent feature vector. This latent feature vector is then used as the vector to be transmitted over the channel.

[0129] Step 6.2 Channel Construction

[0130] A channel model is constructed between the encoder and decoder to simulate noise and interference in real-world communication environments. The channel model is an additive white Gaussian noise (AWGN) channel. By introducing the channel model during training, the network can adapt to different communication environments. The additive white Gaussian noise (AWGN) channel is...

[0131] y = hx + n

[0132] y represents the received signal, i.e., the signal received by the receiver after transmission through the channel. x represents the latent eigenvector, i.e., the original signal transmitted by the transmitter. h represents the channel gain, indicating the attenuation or amplification effect of the channel on the signal. In practical communication, n represents noise.

[0133] Step 6.3 Decoder Construction

[0134] The encoder and decoder are designed to start from the received signal and gradually convert the received signal into data suitable for reconstruction. The received signal first passes through the latent space decoding layer to obtain a feature representation suitable for decoding. The deconvolution layer gradually restores the data of the feature representation in the previous step to obtain the deconvolution result. The deconvolution result is used to generate the reconstructed data through the output layer.

[0135] The overall encoder-decoder channel model is as follows:

[0136]

[0137] s*: Represents the image reconstructed at the receiving end.

[0138] D θ : Represents the decoder, with parameter θ, used to decode the encoded signal back into an image.

[0139] ω: Represents a function used to process the latent variable z.

[0140] z: represents encoder E φ The extracted latent features have a parameter of φ.

[0141] E φ : Represents the encoder, with parameter φ, used to extract latent features from the original image s. E φ The task is to extract latent features from the original image. s passes through E φ The encoder extracts the source data into latent features through a multi-layered neural network. These features contain the core semantic information of the data. The semantic information is presented as a continuous distribution during the initial extraction process; z passes through channel ω, is influenced by ω, and is received by the receiver. The receiver decodes the encoded signal back to the source data. The decoder gradually reconstructs the data through a multi-layered neural network. Each layer performs a non-linear transformation on the input to approximate the representation s* of the source data.

[0142] Step 7: Use dataset 2 from step 5 as training data for the semantic communication network to train the semantic communication network in step 6.

[0143] Step 7.1 Parameter Initialization

[0144] Before training the semantic communication network, the network parameters are initialized. These parameters include those of the encoder, decoder, and channel model. Initialization methods can include random initialization or pre-trained model parameters.

[0145] Step 7.2 Definition of Loss Function

[0146] The network's loss function is defined, using Cross-Entropy Loss (CE) to measure the difference between predicted and true values. For semantic communication networks without hidden information, CE is trained using an autoregressive approach with source data. For semantic communication networks containing hidden information, CE with both source data and labels is trained as the loss function to identify the hidden information.

[0147] CE formula is

[0148]

[0149] H(P): Represents information entropy, used to measure the uncertainty of a random variable X. The higher the information entropy, the higher the uncertainty of the random variable.

[0150] P(X=x i ): represents the probability that a random variable X takes the value xi.

[0151] m: The total number of all possible values ​​that the random variable X can take.

[0152] Step 7.3 Backpropagation

[0153] Dataset 2 is used as input for forward propagation computation, while the gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm. This process starts from the output layer of the decoder and proceeds layer by layer forward until the input layer of the encoder.

[0154] Step 7.4 Parameter Update and Optimizer

[0155] The Adam optimizer is used to update the network parameters of the semantic communication network and the hidden communication classification network based on the calculated gradients.

[0156] Step 7.5 Iterative Training

[0157] The process of forward propagation, loss calculation, backpropagation, and parameter update is repeated until the network performance reaches a predetermined threshold or the training reaches a predetermined number of iterations. In each iteration, the network gradually learns better parameters, improves its semantic understanding of the source data, and the training hyperparameters are modified multiple times to improve the training results. After completing multiple iterations of training, a dedicated semantic communication network is obtained.

[0158] Step 8: Distribute the covert communication classification network and the semantic communication network.

[0159] 8.1 Public Distribution of Semantic Communication Networks

[0160] The public distribution of the Semantic Communication Network will be carried out through multiple channels.

[0161] 8.2 Covert Communication Classification and Distribution

[0162] The clandestine distribution of the classified communications network is strictly limited to partners, ensuring that only authorized entities can access and use the network.

[0163] A public semantic communication network is available for download and use by any user. A covert communication classification network is secretly distributed to collaborating parties. During use, the secret holder sends information, and non-collaborating parties can obtain the transmitted content, but due to the lack of a covert information classifier, they cannot access the transmitted covert information.

[0164] 8.3 In a covert communication system, the sender encodes the pre-transmitted covert information and performs interpolation to obtain a transmission sequence, which is then transmitted over a public channel. After transmission, the receiver in the covert communication system decodes the information to obtain the covert information.

[0165] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cognitively concealed communication method, characterized in that: Includes the following steps, Step 1: Construct dataset 1; The specific implementation method for step one is as follows: The base dataset uses a temporally ordered image data stream, with each image annotated with a sequential label; in addition, Gaussian smoothing is used to construct the training and test sets; the labels only contain the sequence number of the current image; Based on the basic dataset, hidden information is added at the data preparation level by modifying the labels of pre-specified images, thus constructing dataset 1 containing hidden information; Step 2: Using the optical flow method, an optical flow output image is obtained through the optical flow field; Step 3: Train the image interpolation neural network to obtain a trained deep learning method; Step 4: Input the data from Step 1 into the data from Step 3 to obtain the deep learning output image; Step 5: Put the image interpolation data into dataset 1 from step 1 to obtain dataset 2, which will be used as training data for the semantic communication network to train the semantic communication network. The specific implementation method for step five is as follows: The hidden information is encoded according to the label values ​​in dataset 1 to obtain the image stream after the hidden information is encoded; the optical flow output image obtained in step 2 or the neural network output image obtained in step 4 is inserted into the image stream in step 5. Insert several images into each pair of adjacent images in the image stream; obtain the image interpolation data; The image interpolation data is put into dataset 1 to obtain a new dataset, which is used as the training data for the semantic communication network to train the semantic communication network. Step 6: Construct a semantic communication network; Step 6.1: Encoder construction; Step 6.2: Channel Construction; Step 6.3: Decoder Construction; Step 7: Use dataset 2 from step 5 as training data for the semantic communication network to train the semantic communication network from step 6. Step 7.1: Parameter initialization; Step 7.2: Define the loss function; Step 7.3: Backpropagation step; Step 7.4: Parameter Update and Optimizer Steps; Step 7.5: Iterative training; Step 8: Distribute the covert communication classification network and semantic communication network, and realize cognitive covert communication based on the covert communication classification network and semantic communication network; The specific implementation method for step eight is as follows: Step 8.1: The public distribution of the Semantic Communication Network is carried out through multiple channels; Step 8.2: The covert communication classification network is strictly distributed to partners to ensure that only authorized entities can access and use the network; Step 8.3: The sender of the covert communication system encodes the pre-transmitted covert information and performs interpolation to obtain the transmission sequence, which is then transmitted on a public channel. After transmission, the receiver of the covert communication system decodes the information to obtain the covert information.

2. The method as described in claim 1, characterized in that: The specific implementation method for step two is as follows: Based on dataset 1 obtained in step one, the optical flow method is used to obtain the optical flow output image through the optical flow field.

3. The method as described in claim 1, characterized in that: The specific implementation method for step three is as follows: Input dataset 1 contains two adjacent RGB images. and ,in , , Image height, The width is 3, and the number of RGB channels is 3. Perform a PixelUnshuffle downsampling operation on each frame to convert the spatial dimension to the channel dimension and obtain the downsampled features. and Its dimensions are , The scaling factor is [value]; stitching along the channel dimension. and Generate merged features ; pass Convolutional layers will merge features The number of channels is compressed to ,get initial features ;Will Input is sent to a residual group (ResGroup), which contains multiple residual channel attention blocks (RCA Blocks). Each RCA Block performs the following operations: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Perform twice Convolution and ReLU activation generate intermediate features; channel attention weighting is applied to the intermediate features to obtain weighted features; the weighted features are then combined with... The final features are output by summing the residual connections. The method for obtaining weighted features by performing channel attention weighting is as follows: global average pooling is used to generate channel feature descriptors. ,in The number of channels; channel attention weights are generated through two fully connected layers. The formula is: in , For learnable weight matrix, The compression ratio is... The Sigmoid function; combines intermediate features with... Multiply each channel sequentially to obtain the weighted feature; Perform final features Convolution adjusts the number of channels; the final feature channel number is converted using the PixelShuffle upsampling operation. , and according to Block reorganization spatial dimensions to generate intermediate frames ; Calculate intermediate frames With real intermediate frames Pixel reconstruction loss: ,in These are actual intermediate frames; Calculate the perceptual loss based on the pre-trained network and extract feature differences: ,in For the feature extraction function of the pre-trained network; Jointly optimize the total loss function: Iterative optimization to The requirements are met, and the training is completed to obtain a well-trained neural network.

4. The method as described in claim 1, characterized in that: The specific implementation method for step four is as follows: Input the dataset 1 from step one into the neural network trained in step three to obtain the output image of the neural network.

5. The method as described in claim 1, characterized in that: The specific implementation method for step six is ​​as follows: Step 6.1: Construct the encoder; the encoder encodes the source data into a format suitable for transmission, and the decoder decodes the received signal back into the source data; the input layer of the encoder receives the source data, and the input convolutional layer is used to extract features from the source data; Use Transformer layers to capture long-range dependency features in the features extracted by convolutional layers; The latent space mapping layer maps the long-distance dependency features extracted by the Transformer layer to the latent space, obtaining the latent feature vector; potential The feature vector is the vector that will be transmitted over the channel; Step 6.2: Construct a channel model between the encoder and decoder to simulate noise and interference in a real communication environment; The channel model is an additive white Gaussian noise (AWGN) channel; by introducing the channel model during training, the network can adapt to different communication environments; the additive white Gaussian noise (AWGN) channel is... y is the received signal, that is, the signal received by the receiver after transmission through the channel; x is the latent feature vector, that is, the original signal sent by the transmitter; h is the channel gain, which represents the attenuation or amplification effect of the channel on the signal; in actual communication, n is noise. Step 6.3: Build the encoder; The decoder starts by receiving the signal and gradually converts it into data suitable for reconstruction. The received signal first passes through the latent space decoding layer to obtain a feature representation suitable for decoding. The deconvolution layer gradually restores the data of the feature representation in the previous step to obtain the deconvolution result. The deconvolution result is then passed through the output layer to generate the reconstructed data. The overall encoder-decoder channel model is as follows: : Represents the image reconstructed at the receiving end; : Indicates the decoder, with parameters as follows This is used to decode the encoded signal back into an image; : Represents a function used to process the latent variables. ; : Indicates encoder The extracted latent features have parameter ϕ; : Represents the encoder, with parameter ϕ, used to extract data from the original image. Extract latent features from; The task is to extract latent features from the original image; go through The extraction encoder transforms source data into latent features using a multi-layer neural network. These features contain the core semantic information of the data; the semantic information is continuously distributed during the initial extraction process. via channel The signal is received by the receiving end, which then decodes the encoded signal back to the source data. The decoder reconstructs the data step by step through a multi-layer neural network. Each layer of the network performs a non-linear transformation on the input to approximate a representation of the source data. .

6. The method as described in claim 1, characterized in that: The specific implementation method for step seven is as follows: Step 7.1: Before training the semantic communication network, initialize the network parameters; the parameters include the parameters of the encoder, decoder and channel model; the initialization method can be random initialization or initialization using the parameters of the pre-trained model; Step 7.2: Define the loss function of the network and use cross-entropy loss (CE) to measure the difference between the predicted value and the true value; use dataset 2 as training and validation data, and use source data autoregression for the semantic communication network without hidden information, and use CE as the loss to train the semantic communication network; use source data and labels for the semantic communication network with hidden information, and use CE as the loss to train the classification network to identify hidden information. CE is represented as : Represents information entropy, used to measure the uncertainty of a random variable X; the higher the information entropy, the higher the uncertainty of the random variable. : Represents the probability that a random variable X takes the value xi; m: The total number of all possible values ​​of the random variable X; Step 7.3: Forward propagation calculation of dataset 2, and at the same time, the gradient of the loss function with respect to the network parameters is calculated through the backpropagation algorithm; the process starts from the output layer of the decoder and calculates layer by layer forward until the input layer of the encoder; Step 7.4: Update the network parameters of the semantic communication network and the hidden communication classification network based on the calculated gradients using the Adam optimizer; Step 7.5: Repeat the forward propagation, loss calculation, backpropagation and parameter update process until the network performance reaches the predetermined threshold or the training reaches the predetermined number of iterations; In each iteration, the network gradually learns better parameters, improves its semantic understanding of the source data, and modifies the training hyperparameters multiple times to improve the training results. After multiple iterations of training, a dedicated semantic communication network is obtained; Through steps 7.1 to 7.5, the trained semantic communication network can achieve efficient and robust semantic information transmission in complex communication environments.

Images