Anti-noise semantic communication method based on supervised discriminant learning

Abstract

The invention relates to an anti-noise semantic communication method based on supervised discriminant learning, and belongs to the technical field of communication. For the problems of semantic noise and channel noise in image transmission, a deep supervision network is introduced at a semantic task execution end, a supervision layer is additionally arranged behind a hidden layer, and high-quality middle layer features are extracted, so that semantic feature representation is learned; a composite loss function containing terminal loss and adjoint loss is designed, the adjoint loss is introduced into a prototype consistent loss function, and semantic noise influence is reduced through maximum separation constraint; according to the method, the priority of task execution and data recovery is adjusted according to the SNR value so as to adapt to different communication conditions; furthermore, semantic noise is generated by adopting a white box attack method, and the anti-noise performance is enhanced by utilizing a supervision layer output result on the premise of not changing the structure of the original deep learning model; a loss function is designed from the semantic feature level, the anti-noise performance is further enhanced, and data recovery and task execution are both considered.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a noise-resistant semantic communication method based on supervised discriminant learning. Background Art

[0002] With the booming development of machine learning and computer hardware, semantic communication, as a new intelligent communication mode, has attracted widespread attention in recent years. Semantic communication represents a revolutionary advancement in communication technology. Semantic communication systems use deep learning models to extract semantic information from input data and generate feature vectors. The system transmits these feature vectors instead of raw bit sequences, shifting the focus from simple information transmission to the meaning and context of information. Unlike traditional communication systems, semantic communication focuses more on conveying the meaning of information rather than simply symbol transmission. Its key lies in maximizing the retention and transmission of the semantic content of information rather than focusing solely on the accurate delivery of symbols. Although semantic communication can improve communication efficiency by reducing the amount of data transmitted, the introduction of neural networks brings new threats. Unlike traditional communication, semantic communication systems are susceptible to two different types of noise. The first type is physical channel noise, which is found in wireless communications and is caused by physical channel losses. It should be noted that errors caused by channel propagation usually occur before channel decoding. The second type of noise is semantic noise, which causes misunderstanding of semantic information, resulting in a mismatch of semantic information between the transmitter and receiver during semantic communication, and introduces semantic noise at the source and wireless channel. The purpose of semantic noise is to mislead deep learning models. It can easily fool deep learning models by adding subtle perturbations to samples without being detected by humans. It may be added by malicious interferers, or it may be caused by human errors in sampling when collected in nature, resulting in semantic noise in the sampled data. In semantic communication, this noise may cause the receiver to misunderstand or incorrectly decode the transmitted information, affecting the reliability and security of communication. Although there are many methods for generating adversarial perturbations in the field of image processing, studying how to identify and counter semantic noise in wireless communication environments to ensure that the system can accurately decode and reconstruct the transmitted semantic information is an important challenge currently faced by deep learning in the field of semantic communication. Semantic noise usually includes text semantic noise and image semantic noise. Compared with text semantic noise, image semantic noise is generated by adding subtle perturbations to the image, which is almost imperceptible to the human eye, but it is easy to cause a significant decrease in the performance of image machine tasks such as classification and segmentation. Therefore, there is an urgent need for an optimized semantic communication method to provide a guarantee for countering the interference caused by semantic noise. Summary of the invention

[0003] In order to achieve the above object, the present invention provides the following technical solutions:

[0004] A noise-resistant semantic communication method based on supervised discriminant learning comprises the following steps:

[0005] S1. Introducing semantic noise and adopting joint encoder-decoder in semantic communication system, and adding deep supervision network on the semantic task side;

[0006] S2. Generate adversarial samples through Projected Gradient Descent (PGD) attack to interfere with the transmitter, and use iterative algorithm to generate and optimize semantic noise during data transmission;

[0007] S3. Add a deep supervision network to the task end of the semantic communication system, so that each intermediate layer can be directly supervised and semantic noise can be gradually filtered out;

[0008] S4. Design a loss function consisting of terminal loss and adjoint loss to enhance the ability of deep learning models to distinguish samples contaminated by semantic noise;

[0009] S5. Perform semantic noise adversarial training for discriminative learning to enhance the deep learning model’s ability to resist semantic noise.

[0010] Further, the S1 comprises the following steps:

[0011] S11. Adding semantic noise to the data set at the transmitting end of the semantic communication system, and introducing semantic noise in the process of data transmission through the channel;

[0012] S12. The transmitting and receiving ends of the semantic communication system both use a joint source-channel codec to implement the encoding and decoding process;

[0013] S13. The semantic communication system introduces a deep supervision network at the semantic task execution end.

[0014] Further, the S2 comprises the following steps:

[0015] S21. Introducing the Projected Gradient Descent (PGD) attack method, the adversarial sample generated in the t+1th iteration is expressed as:

[0016]

[0017] Among them, x t+1 represents the adversarial sample generated in the t+1th iteration, s represents the original image, and Π x Indicates that the perturbation is limited to a limited range, z represents the semantic task related to the input information, represents the loss function of the deep learning model, Δs represents the attack of semantic noise suffered during the communication process;

[0018] S22. The specific modeling of semantic noise interference during data transmission is as follows:

[0019]

[0020] stΔx≤δ

[0021] Among them, x represents the adversarial sample, Δx represents the semantic noise, and g θ represents the trainable decoder parameters, H represents the channel state information, n represents the channel noise, and H s represents the channel information of semantic noise passing through the channel, and δ represents the power constraint.

[0022] Further, the S3 comprises the following steps:

[0023] S31. Using ResNet (Residual Network) with a residual module as the backbone of the deep supervision network;

[0024] S32. A supervision layer is added after each hidden layer of the deep supervision network, with the feature map as input, and the features are expressed through the fully connected layer to achieve layer-by-layer supervision.

[0025] Further, the S4 comprises the following steps:

[0026] S41. Design a loss function consisting of terminal and accompanying losses, expressed as:

[0027]

[0028] Among them, loss m Indicates terminal loss, loss a represents the accompanying loss, and β represents the weight of the accompanying loss;

[0029] S42. The prototype consistent loss function introduced by the adjoint loss is expressed as:

[0030]

[0031] Among them, y represents the category label of the image, represents the features extracted from i images, Related to the parameters of the codec, c represents the class centroid, y i represents the category label of the i-th graph, represents the category centroid corresponding to the i-th graphic category label, c j Represents the category centroid under the j-th graphic category label;

[0032] S43. Terminal loss, expressed as:

[0033]

[0034] in, represents the loss function for data reconstruction in a semantic communication system, represents the semantic task loss function of the system, and α represents the scaling The value range is adapted to is a hyperparameter, and λ represents a hyperparameter that balances semantic tasks and data recovery.

[0035] Further, the λ is expressed as:

[0036]

[0037] Among them, SNR stands for Signal-to-Noise Ratio, which is used as a measurement index of the system communication environment.

[0038] Furthermore, the category centroid c is obtained by the average value of the feature values ​​learned in each round of the deep learning model.

[0039] Further, the S5 comprises the following steps:

[0040] Step 5.1: Initialize the learning rate l, semantic noise, power constraint δ, and trainable encoder parameters f of the semantic communication system θ , the semantic communication system can train the decoder parameters g θ , the semantic communication system can train the classifier parameter z θ and the trainable supervision layer parameters s in the system θ ;

[0041] Step 5.2, the training data polluted by semantic noise at the source end is encoded by the encoder and transmitted through the channel;

[0042] Step 5.3, calculate and add the semantic noise in the transmission process;

[0043] Step 5.4: pass the information decoded by the decoder to the classifier to perform the task;

[0044] Step 5.5, calculate the terminal loss and the accompanying loss, which are used to calculate the total loss function;

[0045] Step 5.6, calculate the gradient to update the encoder parameters and decoder parameters;

[0046] Step 5.7, calculate the gradient again to update the supervision layer parameters, and calculate the gradient to update the classifier parameters;

[0047] Step 5.8: Calculate the average value of the features extracted by the classifier in the current round and update the centroid.

[0048] The beneficial effects of the present invention are:

[0049] By adding additional supervision signals to the network structure, each intermediate layer can obtain direct supervision feedback during the learning process; semantic noise interference is gradually filtered out when extracting feature information, effectively guiding the network to learn feature representations with more discriminative ability and task relevance; the deep supervision network introduces a supervision layer after the hidden layer, takes the feature map obtained in each layer as the input of the supervision layer, and the supervision layer expresses the features learned by the hidden layer through the fully connected layer, thereby providing a direct supervision signal; the network can not only supervise the final output layer, but also perform intermediate supervision in each hidden layer, so that the entire network can learn features useful for the task at all levels; the layer-by-layer supervision training method not only enhances the expression ability of the network, but also helps the network to gradually filter out unnecessary information and noise from the input data during the learning process; at the same time, a new loss function is designed from the perspective of semantic similarity, and the semantic noise is moved away from the centroid of the semantic feature by introducing a prototype consistency loss function combined with the clustering idea, and a new loss is designed to take into account both data recovery and task execution; the invention proposes a noise-resistant semantic communication method based on supervised discriminative learning, which comprehensively considers various business needs and can be used by the system, thereby counteracting the interference caused by semantic noise to the semantic communication system.

[0050] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0052] Figure 1 It is a flow chart of a noise-resistant semantic communication method based on supervised discriminant learning according to an embodiment of the present invention;

[0053] Figure 2 An image with semantic noise added in the semantic communication system of an embodiment of the present invention;

[0054] Figure 3 A schematic diagram of the network hierarchy of the semantic task end of an embodiment of the present invention;

[0055] Figure 4 This is a flowchart of semantic noise adversarial training for discriminative learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0057] See also Figure 1 , is a flow chart of a noise-resistant semantic communication method based on supervised discriminant learning according to an embodiment of the present invention; the noise-resistant semantic communication method based on supervised discriminant learning according to an embodiment of the present invention mainly comprises the following steps:

[0058] Step 1: Design of semantic communication scenarios for image transmission subject to semantic noise interference:

[0059] In the semantic communication system for image transmission, a noise-resistant semantic communication method based on supervised discriminant learning is studied under the interference of semantic noise that is imperceptible to the human eye. This scenario is mainly aimed at image transmission. Since image semantic noise is generated by adding subtle perturbations to the image, it is almost imperceptible to the human eye, but it is easy to cause a significant decrease in the performance of image machine tasks (classification, segmentation, etc.). Therefore, a semantic communication system for image transmission under the interference of semantic noise is designed.

[0060] Step 1 specifically includes the following steps:

[0061] Step 1.1: Introduction of semantic noise,In the semantic communication system, in order to simulate and deal with the semantic interference that may be encountered in the actual communication process, semantic noise is deliberately added to the data set at the sender. At the same time, additional semantic noise is also introduced in the process of data transmission through the channel to fully simulate the complex situation in the communication environment.

[0062] Step 1.2: Codec selection,In order to improve the efficiency and accuracy of the communication system, the semantic communication system uses a joint source-channel codec at both the sender and the receiver.,This codec can simultaneously consider the information characteristics of the source and the transmission characteristics of the channel,,thus achieving a more efficient encoding and decoding process.

[0063] Step 1.3: Introduction of deep supervision network. Different from the existing semantic communication framework, this semantic communication system innovatively introduces a deep supervision network at the semantic task execution end. This network can extract higher-quality intermediate layer features, which are important for resisting the interference of semantic noise. With the introduction of the deep supervision network, the system can more effectively process and understand the received information, thereby improving the accuracy and reliability of communication.

[0064] Step 2: Generation of semantic noise:

[0065] See also Figure 2 , is an image with semantic noise added in the semantic communication system of an embodiment of the present invention. The purpose of semantic noise is to mislead the deep learning model, and it can easily deceive the deep learning model by adding subtle disturbances to the sample without being detected by humans. It may be added by malicious interferers, or it may be that the sampled data contains semantic noise due to human errors in sampling when collected in nature.

[0066] Step 2 specifically includes the following steps:

[0067] Step 2.1: The interference of semantic noise on the sending end:

[0068] This semantic noise is generated in the encoding stage and may be caused by a variety of reasons. In order to achieve the target process of semantic noise at the sender, a white box attack method is designed, which uses the gradient information of the deep learning model to generate deceptive input samples, causing the deep learning model to make incorrect predictions. Specifically, the Projected Gradient Descent (PGD) attack method is introduced. It can be expressed as follows:

[0069] x t+1 =Π x+Δs (x t +α·sign(▽ s L(f θ (x t ),z)))

[0070] Among them, x t+1 represents the adversarial sample generated in the t+1th iteration, s represents the original image, and Π x Indicates that the perturbation is limited to a limited range to ensure that the perturbation does not cause the input to exceed a reasonable range. z represents the semantic task related to the input information, represents the loss function of the deep learning model, and Δs represents the attack of semantic noise suffered during the communication process.

[0071] Step 2.2, interference from semantic noise during data transmission: Due to the openness of the wireless channel, semantic noise can be easily introduced into the semantic communication system if there is malicious interference or other factors. Semantic noise may distort the transmitted semantic information, thereby affecting the accuracy of decoding at the receiving end. Therefore, the system must also consider the semantic noise interference that may be suffered during information transmission to ensure that the system can operate effectively in a noisy environment and accurately transmit semantic information. This comprehensive consideration and processing method helps to improve the system's anti-semantic noise performance and reduce information transmission errors caused by semantic noise, thereby ensuring the reliability and stability of semantic communication. The specific modeling is as follows:

[0072]

[0073] stΔx≤δ

[0074] Among them, x represents the adversarial sample, Δx represents the semantic noise, and g θ represents the trainable decoder parameters, H represents the channel state information, n represents the channel noise, and H s represents the channel information of the semantic noise passing through the channel, and δ represents the power constraint, that is, the semantic noise received must not be higher than this threshold. The generation algorithm of the semantic noise received during data transmission specifically includes the following steps:

[0075] Step 2.2.1, initialize the learning rate l, semantic noise Δx, and power constraint δ;

[0076] Step 2.2.2: If the semantic noise Δx is less than the power constraint δ, then the formula is:

[0077]

[0078] Δx k+1 =Δx k +l·σ

[0079] The semantic noise is calculated, and then iterative training is performed to find an optimal solution for the semantic noise.

[0080] Step 3: Introduce the structure of the deep supervision network to obtain the network structure of the semantic task side:

[0081] See also Figure 3: The network hierarchy diagram of the semantic task end of an embodiment of the present invention; the deep supervision network refers to a supervised learning method that adds additional learning methods to the network. Its core advantage is that it can add additional supervision signals to the network structure so that each intermediate layer can obtain direct supervision feedback during the learning process. The system can gradually filter out semantic noise interference when extracting feature information, and effectively guide the network to learn feature representations that are more discriminative and task-relevant. The specific process is that after the decoder decodes the semantic information, this information is passed layer by layer to each layer of the task execution part. At each layer, after the features are extracted by the convolution operation, they are immediately sent to the supervision layer, which processes these features through the fully connected layer and the (RectifiedLinear Unit, ReLU) activation function, and calculates the loss with the target. Therefore, each layer can be adjusted and optimized through direct supervision signals to ensure that the features extracted at each level are useful for the final task.

[0082] Step 3 specifically includes the following steps:

[0083] Step 3.1. Considering the complexity of the task and the performance of the deep learning model, the Residual Network (ResNet) with a residual module is used as the backbone of the designed deep supervision network.

[0084] Step 3.2: The structure of the deep supervision network is to add a supervision layer after each layer:

[0085] The deep supervised network introduces a supervision layer after the hidden layer and uses the feature map obtained in each layer as the input of the supervision layer. The supervision layer expresses the features learned by the hidden layer through the fully connected layer, thereby providing a direct supervision signal. In this way, the network can not only supervise the final output layer, but also perform intermediate supervision in each hidden layer, so that the entire network can learn features that are useful for the task at all levels. This layer-by-layer supervised training method not only enhances the network's expressive power, but also helps the network gradually filter out unnecessary information and noise from the input data during the learning process.

[0086] Step 4: Loss function design:

[0087] First, in order to combat the interference of semantic noise, the adjoint loss of the deep supervision network introduces a prototype consistency loss function. Its main idea is to map the intermediate features of the network into a low-dimensional prototype space so that the feature points of the same category are closer in the prototype space, while the feature points of different categories are more dispersed. The purpose of this approach is to enhance the ability of the deep learning model to distinguish samples contaminated by semantic noise, so that the deep learning model can better maintain the accuracy of classification when facing semantic noise. Secondly, in order to simultaneously consider the execution of semantic tasks and the quality of image reconstruction, the terminal loss is further designed.

[0088] Step 4 specifically includes the following steps:

[0089] Step 4.1: In order to combat the interference of semantic noise while taking into account both the task and image restoration, a loss function consisting of terminal and accompanying losses is designed. The loss function is specifically expressed as:

[0090]

[0091] Among them, loss m represents the terminal loss, β represents the weight of the accompanying loss, loss a Indicates incidental loss.

[0092] Step 4.2: The prototype consistent loss function introduced by the adjoint loss is expressed as:

[0093]

[0094] Among them, y represents the category label of the image, represents the features extracted from the i-th image. This parameter is related to the parameters of the codec. c represents the centroid of the category. y i represents the category label of the i-th graph, represents the category centroid corresponding to the i-th graphic category label, c j represents the category centroid under the j-th graphic category label; where y represents the category label of the image, represents the features extracted from the i images, c represents the category centroid, and the category centroid is obtained by the average value of the feature values ​​learned in each round of the deep learning model.

[0095] The training process of the prototype consistency loss function includes four steps: feature extraction, prototype generation, feature mapping, and category separation. First, a deep neural network is used to extract features from the input samples. Secondly, the prototype vector of each category is calculated, which is the average of the features of all samples in that category. Then, the extracted features are mapped to the prototype space to form a representation of the sample in that space. Finally, by minimizing the distance between samples of the same category and the prototype and maximizing the distance between samples of different categories and the prototype, category separation is achieved, enhancing the robustness of the deep learning model to semantic noise.

[0096] Step 4.3: In order to consider both the execution of the semantic task and the quality of image reconstruction, the terminal loss is specifically expressed as:

[0097]

[0098] in, represents the loss function for data reconstruction in a semantic communication system, It represents the semantic task loss function of the system. α is used to scale The value range is adapted to λ is a hyperparameter used to balance the semantic task and data recovery.

[0099] Since λ is a hyperparameter used to weigh the semantic task and data recovery, but the weight ratio between data recovery and semantic task is unclear. Therefore, how to set the hyperparameter λ is very important. According to previous machine learning experience, the specific situation of the hyperparameter can only be adjusted manually. In order to set the hyperparameter more conveniently, a flexible hyperparameter processing method is designed. Specifically, considering the impact of semantic noise interference on the system, the signal-to-noise ratio (SNR) is used as a measurement index of the system communication environment. If the communication conditions are poor, that is, the SNR value is small, the system gives priority to the task execution rate and focuses on the semantic task. If the communication conditions are good, that is, the SNR value is large, the system will consider the task execution while also allocating part of the performance to consider the data recovery situation. λ is specifically expressed as:

[0100]

[0101] Step 5: Adversarial training for discriminative learning:

[0102] Since the sensitivity of prototype selection is the main limitation of the prototype-consistent loss function, it is easily affected by the distribution of training data and semantic noise. In order to overcome this limitation, it is proposed to combine adversarial training and expand the training set by introducing adversarial samples to make the deep learning model more robust in the face of interference and noise. Adversarial training uses samples contaminated by semantic noise as more challenging examples to help the deep learning model learn more comprehensive and robust feature representations and decision boundaries, thereby improving the effect of the prototype-consistent loss function. The main idea of ​​semantic noise adversarial training is to add samples contaminated by semantic noise to the training set, so as to enhance the deep learning model's ability to resist semantic noise. Adversarial training is an important technology for improving the robustness of deep learning models. In this way, the trained deep learning model can not only work well under normal conditions, but also maintain a high accuracy rate when facing semantic noise interference, thereby enhancing the overall robustness of the deep learning model. Specifically, adversarial training generates adversarial samples so that the deep learning model is constantly exposed to challenging noise environments during the training process, thereby improving its performance in practical applications. See Figure 4 , which is a flowchart of semantic noise adversarial training for discriminative learning in an embodiment of the present invention.

[0103] Step 5 specifically includes the following steps:

[0104] Step 5.1: Initialize the learning rate l, semantic noise, power constraint δ, and trainable encoder parameters f of the semantic communication system θ , the semantic communication system can train the decoder parameters g θ , the semantic communication system can train the classifier parameter z θ and the trainable supervision layer parameters s in the system θ ;

[0105] Step 5.2, the training data polluted by semantic noise at the source end is encoded by the encoder and transmitted through the channel;

[0106] Step 5.3, calculate and add the semantic noise in the transmission process;

[0107] Step 5.4: pass the information decoded by the decoder to the classifier to perform the task;

[0108] Step 5.5, calculate the terminal loss and the accompanying loss, which are used to calculate the total loss function;

[0109] Step 5.6, calculate the gradient to update the encoder parameters and decoder parameters;

[0110] Step 5.7, calculate the gradient again to update the supervision layer parameters, and calculate the gradient to update the classifier parameters;

[0111] Step 5.8: Calculate the average value of the features extracted by the classifier in the current round and update the centroid.

[0112] The noise-resistant semantic communication method based on supervised discriminant learning proposed in this embodiment is a semantic communication system method for resisting semantic noise and channel noise for image transmission. The method takes into account the noise interference received at the sending end and during data transmission. In order to resist the interference of semantic noise, a deep supervision network is designed and introduced at the semantic task execution end. By introducing a supervision layer after the hidden layer, the feature map obtained in each layer is used as the input of the supervision layer to obtain high-discrimination and high-quality intermediate layer features, so that the semantic communication system can learn more robust and meaningful feature representations related to the task. At the same time, in order to counter the interference of semantic noise while taking into account both tasks and image restoration, a loss function consisting of terminal and accompanying losses is designed. Among them, the terminal loss considers the execution of semantic tasks and the quality of image restoration at the same time, and the accompanying loss further improves the noise resistance of the system. In order to better resist semantic noise interference, the accompanying loss of the deep supervision network introduces a prototype consistency loss function, which can effectively reduce the impact of semantic noise by including a simple maximum separation constraint in the target of the deep network. However, the traditional cross entropy loss function does not have such a function, which makes the resulting deep learning model weak in resisting semantic noise. This loss function can be understood as: defining convex category-specific classification regions for each category, where any samples outside all these regions are considered noise. It is used to learn feature representations between samples with similar semantics and encourage feature representations between different categories to separate from each other. The goal of this loss function is to cluster samples of the same category in similar positions in the feature space, while dispersing samples of different categories. In order to simultaneously consider the execution of semantic tasks and the quality of image reconstruction, and the impact of noise interference on the system, SNR is used as a measurement index of the system communication environment in this paper. When the communication conditions are poor, that is, the SNR value is small, the system prioritizes the task execution rate and focuses on the semantic task. When the communication conditions are good, that is, the SNR value is large, the system considers task execution while also allocating part of the performance to consider data recovery.

[0113] The present invention proposes a noise-resistant semantic communication method based on supervised discriminant learning, which is a method starting from the semantic feature level, adding a deep supervision network at the semantic task end, designing a loss function and considering data recovery and task execution at the same time. First, since semantic noise is a kind of noise that cannot be recognized by the human eye, a white box attack method is used to generate semantic noise and add it to the sending end and transmission process of the system. Secondly, in order to better combat the interference of semantic noise, by adding a deep supervision network at the semantic task end, that is, the structure is to add a supervision layer after each layer. This method does not modify the original deep learning model structure, but only uses the output result of the supervision layer after the ReLU activation function. Finally, a loss function is designed based on semantic similarity to further enhance the anti-noise performance, while also taking into account data recovery and task execution.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A noise-resistant semantic communication method based on supervised discriminant learning, characterized in that: The following steps are involved: S1. Introducing semantic noise and adopting joint encoder-decoder in semantic communication system, and adding deep supervision network on the semantic task side; S2. Generate adversarial samples through Projected Gradient Descent (PGD) attack to interfere with the transmitter, and use iterative algorithm to generate and optimize semantic noise during data transmission; S3. Add a deep supervision network to the task end of the semantic communication system, so that each intermediate layer can be directly supervised and semantic noise can be gradually filtered out; S4. Design a loss function consisting of terminal loss and adjoint loss to enhance the ability of deep learning models to distinguish samples contaminated by semantic noise; S5. Perform semantic noise adversarial training for discriminative learning to enhance the deep learning model’s ability to resist semantic noise.

2. The noise-resistant semantic communication method based on supervised discriminant learning according to claim 1, characterized in that: The S1 comprises the following steps: S11. Adding semantic noise to the data set at the transmitting end of the semantic communication system, and introducing semantic noise in the process of data transmission through the channel; S12. The transmitting and receiving ends of the semantic communication system both use a joint source-channel codec to implement the encoding and decoding process; S13. The semantic communication system introduces a deep supervision network at the semantic task execution end.

3. The noise-resistant semantic communication method based on supervised discriminant learning according to claim 1, characterized in that: The S2 comprises the following steps: S21. Introducing the Projected Gradient Descent (PGD) attack method, the adversarial sample generated in the t+1th iteration is expressed as: Among them, x t+1 represents the adversarial sample generated in the t+1th iteration, s represents the original image, and Π x Indicates that the perturbation is limited to a limited range, z represents the semantic task related to the input information, represents the loss function of the deep learning model, Δs represents the attack of semantic noise suffered during the communication process; S22. The specific modeling of semantic noise interference during data transmission is as follows: stΔx≤δ Among them, x represents the adversarial sample, Δx represents the semantic noise, and g θ represents the trainable decoder parameters, H represents the channel state information, n represents the channel noise, and H s represents the channel information of semantic noise passing through the channel, and δ represents the power constraint.

4. The noise-resistant semantic communication method based on supervised discriminant learning according to claim 1, characterized in that: The S3 comprises the following steps: S31. Using ResNet (Residual Network) with a residual module as the backbone of the deep supervision network; S32. A supervision layer is added after each hidden layer of the deep supervision network, with the feature map as input, and the features are expressed through the fully connected layer to achieve layer-by-layer supervision.

5. The noise-resistant semantic communication method based on supervised discriminant learning according to claim 1, characterized in that: The S4 comprises the following steps: S41. Design a loss function consisting of terminal and accompanying losses, expressed as: Among them, loss m Indicates terminal loss, loss a represents the accompanying loss, and β represents the weight of the accompanying loss; S42. The prototype consistent loss function introduced by the adjoint loss is expressed as: Among them, y represents the category label of the image, represents the features extracted from i images, Related to the parameters of the codec, c represents the class centroid, y i represents the category label of the i-th graph, represents the category centroid corresponding to the i-th graphic category label, c j Represents the category centroid under the j-th graphic category label; S43. Terminal loss, expressed as: in, represents the loss function for data reconstruction in a semantic communication system, represents the semantic task loss function of the system, and α represents the scaling The value range is adapted to is a hyperparameter, and λ represents a hyperparameter that balances semantic tasks and data recovery.

6. The noise-resistant semantic communication method based on supervised discriminant learning according to claim 5, characterized in that: The λ is expressed as: Among them, SNR stands for Signal-to-Noise Ratio, which is used as a measurement index of the system communication environment.

7. The noise-resistant semantic communication method based on supervised discriminant learning according to claim 5, characterized in that: The category centroid c is obtained by the average value of the feature values ​​learned in each round of the deep learning model.

8. The noise-resistant semantic communication method based on supervised discriminant learning according to claim 1, characterized in that: The S5 comprises the following steps: Step 5.1: Initialize the learning rate l, semantic noise, power constraint δ, and trainable encoder parameters f of the semantic communication system θ , the semantic communication system can train the decoder parameters g θ , the semantic communication system can train the classifier parameter z θ and the trainable supervision layer parameters s in the system θ ; Step 5.2, the training data polluted by semantic noise at the source end is encoded by the encoder and transmitted through the channel; Step 5.3, calculate and add the semantic noise in the transmission process; Step 5.4: pass the information decoded by the decoder to the classifier to perform the task; Step 5.5, calculate the terminal loss and the accompanying loss, which are used to calculate the total loss function; Step 5.6, calculate the gradient to update the encoder parameters and decoder parameters; Step 5.7, calculate the gradient again to update the supervision layer parameters, and calculate the gradient to update the classifier parameters; Step 5.8: Calculate the average value of the features extracted by the classifier in the current round and update the centroid.

Images