Personalized semantic communication method based on knowledge distillation
By employing a two-stage training method based on knowledge distillation, semantic and channel coding are separated, solving the problems of channel simulation and heterogeneous data in federated learning, and realizing efficient and low-latency transmission of personalized semantic communication.
Patent Information
- Application Number
- CN202511554923.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-29
- Publication Date
- 2026-01-27
AI Technical Summary
How to achieve personalized semantic communication while protecting user privacy, and solve the problems of channel simulation, heterogeneous data and communication burden, especially by integrating channel coding modules into federated learning frameworks, handling non-independent and identically distributed and labeled heterogeneous data, and reducing model transmission latency and resource consumption.
A two-stage training method based on knowledge distillation is adopted. In the first stage, the semantic encoder is trained on the local client through federated learning. In the second stage, a channel coding module is added during centralized training to separate the semantic and channel coding processes and reduce model scaling and communication resource consumption by utilizing knowledge distillation.
It achieves personalized semantic communication while protecting user privacy, reduces communication resource consumption and latency in model aggregation, adapts to changes in the communication environment without retraining federated learning, and solves the problems of channel simulation and heterogeneous data.
Smart Images
Figure CN121413627A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of semantic communication technology, specifically relating to a personalized semantic communication method based on knowledge distillation. Background Technology
[0002] Semantic communication utilizes deep learning models to extract abstract semantic information highly relevant to the task objective from raw data. At the receiving end, this semantic information can be directly used for task execution without reconstructing the original data, thus significantly improving task execution efficiency. Semantic communication effectively removes redundant data irrelevant to the task, greatly improving data compression rates and is a key technology for breaking the Shannon limit.
[0003] Knowledge distillation involves two models with identical architectures but different numbers of parameters. The larger model, called the teacher model, is fully trained and exhibits superior performance. The smaller model, called the student model, has far fewer parameters but maintains the same architecture; it is an untrained model. Training data is simultaneously fed into both the teacher and student models. The student model undergoes supervised learning using the teacher model's output as a label, ultimately resulting in a converged student model. This approach not only compresses model scale while maintaining performance but also significantly accelerates convergence.
[0004] Federated learning is a distributed model training framework that includes local and global models, which are identical. The local model is trained on the user's local device and, after multiple iterations, is uploaded to a central server. Multiple local models are aggregated on the server to obtain the global model. The aggregated global model is then downloaded back to the user's device for another round of iterations. This process continues until the global model converges. Federated learning is a key technology for providing personalized intelligent services to users while protecting their privacy.
[0005] Semantic communication can provide users with low-latency, high-efficiency intelligent communication services. However, users expect to receive personalized semantic communication services based on their own habits and preferences. At the same time, users do not agree to upload their personal data to a central server, as this would infringe on their privacy. Therefore, how to achieve privacy-preserving personalized semantic communication remains an unsolved challenge. An effective solution is to use a federated learning framework to train the semantic communication model, but this faces three technical difficulties: 1. Channel Simulation. Semantic communication requires consideration of the channel simulation process for wireless transmission, and this process does not participate in model training and backpropagation. How to integrate the channel coding module into the federated learning framework is a problem that needs to be solved.
[0006] 2. Heterogeneous Data. This includes two types of heterogeneous environments: heterogeneous distribution and heterogeneous labeling. First, user data is usually not independently and identically distributed, which can lead to model drift and make global model convergence difficult. Second, user data is typically only partially labeled, with the majority being unlabeled, posing a significant challenge to traditional supervised learning.
[0007] 3. Communication Burden. Federated learning requires multiple rounds of global model aggregation, necessitating frequent uploads of local models to the central server, which consumes significant communication resources. Furthermore, in pursuit of model performance, most semantic communication schemes employ large model scales, inevitably leading to transmission latency.
[0008] To address the aforementioned technical challenges, existing solutions can be mainly categorized into three types: 1. Channel Simulation. The channel coding module is retained within the semantic communication framework and participates in federated learning, but the channel coding part does not participate in model aggregation. While feasible, this approach has poor reusability. If the communication environment changes, the channel model needs to be modified, and the federated learning process needs to be re-executed to train the entire model, which is costly and increases the computational burden on the client.
[0009] 2. Heterogeneous Data. Generative AI is employed to generate missing data on the client side, compensating for differences in data distribution across different clients. However, this approach struggles to generate data that doesn't exist. Furthermore, data generation relies heavily on labels; if the data has few or no labels, the generated data will be highly homogenized and of low quality.
[0010] 3. Communication burden. Existing solutions no longer transmit the complete model, but rather the gradients of the model updates. The server aggregates the gradient information and then updates the parameters of the global model. However, the magnitude of the gradient is positively correlated with the model scale; larger models also generate a large amount of gradient information, which cannot fundamentally alleviate the consumption of communication resources and transmission latency. Summary of the Invention
[0011] In view of this, the purpose of this invention is to provide a personalized semantic communication method based on knowledge distillation.
[0012] To achieve the above objectives, the present invention provides the following technical solution: A personalized semantic communication method based on knowledge distillation includes the following steps: S1. The local client downloads the teacher model and the local student model. The local client performs knowledge distillation and federated learning, while simultaneously inputting data into the teacher model. and student model In the middle, student model Teacher model The semantic representation is used as a label, and the semantic extraction ability is learned. After distillation, the student models are uploaded to the central server for aggregation to obtain the global model. ,go through After rounds of training, a converged global student model is obtained. Global student model It includes the user's personalized knowledge; S2, Preserve the global student model semantic encoder Furthermore, a channel coding module and a semantic decoder are added to this; let the input image be... The semantic features obtained after passing through the semantic encoder are The channel coding module is used for semantic features Channel coding is performed to obtain the encoded signal that can be transmitted. The encoded signal can be transmitted The lossy semantic signal is sent to the receiving end and received at the receiving end. The receiving end will receive the lossy semantic signal. Convert into semantic information The data is then input into the corresponding semantic decoder to complete the semantic communication task.
[0013] As a further preferred embodiment of the present invention, the local client is... One, represented as Each client Have local dataset and local student models ,in Representing the iteration rounds of federated learning, the dataset is represented as... The number of all training samples is expressed as Teacher Model and student model Each has a pair of encoders and decoders, represented as follows: and .
[0014] As a further preferred embodiment of the present invention, using and The parameter weights of the global student model and the local student model are represented respectively. The aggregation process in S1 is expressed as follows:
[0015] As a further preferred embodiment of the present invention, the channel coding module in S2 performs semantic feature encoding. Channel coding is performed to obtain the encoded signal that can be transmitted. As shown below: Among them, the function Represents the channel coding module, parameters Represents the current signal-to-noise ratio. This represents the encoded signal that can be transmitted. After power normalization, the transmission process is represented as follows: in, Represents the channel gain matrix. Represents channel noise. This represents the lossy semantic signal at the receiving end.
[0016] As a further preferred embodiment of the present invention, in S1, the semantic extraction capability of the teacher model is transferred to the student model through knowledge distillation. The encoders and decoders of both the student and teacher models are composed of Transformer modules. In the student model, the encoder and decoder respectively include... and Each Transformer block, in the teacher model, contains the encoder and decoder respectively. and A Transformer block, and .
[0017] As a further preferred embodiment of the present invention, S1 specifically includes the following operations: Original image Divided into Image patches, represented as: Image mask training method is adopted, based on mask rate A fixed proportion of image patches are discarded, and the remaining image patches are used for model training, represented as... , where set This represents the set of indices of the visual image patches that were not discarded by the mask. This represents the number of image patches that are retained. The inputs are simultaneously fed into both the student model and the teacher model, resulting in semantic vectors for the student model. and teacher model semantic vector The process is represented as: The student model uses the output of the teacher model as labels for supervised learning, using a function. Representing L1 loss, the encoder distillation loss is expressed as: in, Represents a fully connected layer, used for alignment. and The dimension; The learnable mask vectors are inserted into the semantic vectors respectively. and In the process of restoring the image patch order during segmentation, the following is represented: in, Representative and semantic vectors The set of learnable vectors of the same dimension corresponds to image patches discarded during the mask training phase. This represents the set of subscripts for the masked image blocks. and These represent the semantic features of the student model and teacher model after being filled in, respectively, and are input into the decoders of the student model and teacher model. The output of the student model is used... This indicates that the first term in the teacher model decoder is taken. Output of the Layer Transformer module For each label, the decoder distillation loss is calculated and expressed as: Use weights By weighting the distillation loss, the complete distillation loss is expressed as: Based on this loss, the parameters of the student model are updated through backpropagation, executed locally. After rounds of knowledge distillation, the local student model The model will be uploaded to the server for model aggregation to obtain a global model. The local device downloads the global model from the server to update the local student model, thus obtaining the updated model. A new round of federated learning begins, and this process is repeated until the global model converges, resulting in the global student model. .
[0018] As a further preferred embodiment of the present invention, the architecture of the channel coding module in S2 includes: the signal-to-noise ratio (SNR) is expanded into a multi-dimensional vector after passing through a fully connected neural network (MLP), and then converted into channel weight information and semantic features through a sigmoid function. After MLP dimensionality expansion and fusion with channel weights, the fused result is transformed into semantic weights via a Sigmoid function and then combined with the original semantic features. The weighted result is then normalized to obtain the encoded, transmittable signal. .
[0019] As a further preferred embodiment of the present invention, the loss function for semantic transmission adopts the information bottleneck, and is expressed as: in, Represents the loss weight. Represents mutual information, Represents the distribution of image samples and semantic feature distribution mutual information, Represents semantic feature distribution and task tag distribution Mutual information.
[0020] As a further preferred embodiment of the present invention, the image samples collected from the dataset can be represented as follows: The corresponding label samples and semantic feature samples are represented as follows: and The probability distributions of image samples and task labels are estimated by sampling from the dataset and are expressed as follows: and Probability distribution of semantic features It is unknown; variation is introduced through reparameterization techniques. To estimate and obtain mutual information The upper realm : Based on user Local dataset By sampling and estimating the probability distribution, we obtain the following form: in, represent Divergence, estimated numerically Divergence, as shown below: in, For the Sigmoid function, ,in The variance represents the channel noise. Mutual Information Through loss function estimation under specific task scenarios, for classification tasks, this mutual information is equivalent to: in, Represents cross-entropy. A probability vector representing the image category; For the reconstruction task, this mutual information is equivalent to: in, Represents the mean square error. This represents the original image that has been reconstructed.
[0021] The beneficial effects of this invention are as follows: This invention separates the channel coding module from semantic transmission to address the technical challenges of channel simulation. The channel coding module is no longer coupled to the semantic communication system but is an independent module designed according to the communication environment. Therefore, the method comprises two stages. In the first stage, a federated learning architecture is used to train the semantic encoder and decoder in a federated learning manner, focusing on semantic extraction and excluding channel coding. In the second stage, a centralized training strategy is adopted, retaining the semantic encoder from the first stage, adding the channel coding module on top of it, and designing a corresponding semantic decoder according to the specific task. The encoder only undergoes fine-tuning, while other modules are retrained. The advantage of this method is that semantic extraction needs to conform to user preferences; therefore, this invention uses a federated learning approach to collect personalized knowledge. The channel coding module is a function shared by all users and is independent of user preferences, allowing for centralized training. Therefore, this invention separates semantics and channel, with the former using federated learning and the latter using centralized learning. Even if the communication environment changes, only the communication module needs to be modified and the semantic encoder fine-tuned; the federated learning process does not need to be re-executed, thus maximizing reusability. The semantic encoder will store user-personalized knowledge and can flexibly change the channel coding module and semantic decoder according to the communication environment and task requirements, thus solving the technical difficulties of channel simulation.
[0022] This invention employs a knowledge distillation-based pre-training scheme to fully utilize unlabeled data while reducing model scale, thus addressing the technical challenges of heterogeneous data and communication burden. The essence of non-independent and identically distributed (ICD) models is the lack of specific data; simultaneously, heterogeneous labels introduce a large amount of redundant unlabeled data. Therefore, fully utilizing unlabeled data can effectively alleviate model drift. In the first stage, knowledge distillation is used, where the teacher model is fully trained, possessing a larger scale and stronger semantic extraction capabilities, while the student model has far fewer parameters. User data is simultaneously input into both models. The semantic representation extracted from the original data by the teacher model serves as the label, supporting the supervised training process of the student model. The student model participates in federated learning model aggregation; the smaller model scale effectively reduces communication bandwidth consumption. Ultimately, the converged student model possesses accurate semantic extraction capabilities. In the second stage, the semantic encoder of the student model is retained, and a channel coding module and a task-specific decoder are added. Centralized training is employed to endow the pre-trained student model with the ability to perform specific tasks, thereby obtaining a complete semantic communication system. Distillation and pre-training schemes can effectively address the problem of heterogeneous data on clients, independent of user labels and data distribution, thus solving the technical challenges of heterogeneous data. Knowledge distillation can significantly reduce model scale, alleviate the communication resource consumption caused by model aggregation, and solve the technical challenge of communication burden.
[0023] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0024] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart illustrating a personalized semantic communication method based on knowledge distillation according to the present invention. Figure 2 This is a schematic diagram of the knowledge distillation process of the present invention; Figure 3 This is a schematic diagram of the personalized semantic transmission system of the present invention; Figure 4 This is a schematic diagram of the channel coding module of the present invention; Figure 5 This is a flowchart of the present invention. Detailed Implementation
[0025] like Figures 1-5As shown, this invention designs a personalized semantic communication method based on knowledge distillation to achieve intelligent image transmission services. The image at the sending end has its semantic features extracted, and after receiving the semantic features, the receiving end performs two types of tasks: image classification and image reconstruction. The core process of this invention consists of two parts. In the first stage of pre-training, a federated learning strategy is used to protect user privacy, and a semantic encoder is trained through knowledge distillation, enabling the encoder to learn the ability to extract personalized user knowledge. In the second stage of fine-tuning, a channel coding module and a semantic decoder are added, and a centralized scheme is used to activate specific functions of the encoder.
[0026] 1) System Model: Figure 1 A system model of the present invention is shown. Specifically, the scenario includes... A client can be represented as Each client Have local dataset and local model ,in This represents the number of iterations in the federated learning process. The dataset can be represented as... The total number of training samples can be expressed as Teacher Model and student model Each has a pair of encoders and decoders, which can be represented as follows: and .
[0027] In the first phase, the local client performs knowledge distillation and federated learning. Data is simultaneously fed into the teacher model. and student model In this process, the student model uses the semantic representation of the teacher model as labels to learn semantic extraction capabilities. After... After distillation, the student models are uploaded to the central server for aggregation to obtain the global model. .use and Let represent the parameter weights of the global student model and the local student model, respectively. Then, the aggregation process can be represented as: go through After rounds of training, a converged global student model is obtained. It contains the user's personalized knowledge, which can be simplified as In the second phase, retain encoder Furthermore, a channel coding module and a semantic decoder are added to this. Assume the input image is... The semantic features obtained after passing through the semantic encoder are The channel coding module is responsible for channel coding the semantic information, and this process can be represented as: Where the function Represents the channel coding module, parameters Represents the current signal-to-noise ratio. This represents the encoded signal that can be transmitted. After power normalization, the transmission process can be represented as follows: in Represents the channel gain matrix. Represents channel noise. This represents the lossy semantic signal received at the receiver. Subsequently, the receiver will... Convert into semantic information The data is then input into the corresponding decoder to complete the semantic communication task. As shown in Figure 1, two types of decoders are considered here, one for image classification and the other for image reconstruction. The former will demonstrate the system's ability to understand semantics, while the latter will demonstrate the system's usability in data reconstruction.
[0028] 2) Knowledge Distillation: In the first stage, the semantic extraction capabilities of the teacher model are transferred to the student model through knowledge distillation. In this invention, the encoders and decoders of both the student and teacher models are composed of Transformer modules, as shown in Figure 2. In the student model, the encoder and decoder respectively contain... and Each Transformer block, in the teacher model, contains an encoder and a decoder, respectively. and A Transformer block, and .
[0029] Original image Divided into An image patch can be represented as: This invention employs an image mask training scheme based on mask rate. A fixed proportion of image patches will be discarded, and the remaining image patches will be used for model training, which can be represented as follows: , where the set This represents the set of indices of the visual image patches that were not discarded by the mask. This represents the number of image patches that are retained. Simultaneously input into both the student model and the teacher model, semantic vectors are obtained respectively. and This process can be represented as: The student model uses the output of the teacher model as labels for supervised learning. If a function is used... Let L1 represent the loss, then the encoder distillation loss can be expressed as: in Represents a fully connected layer, used for alignment. and Dimensions.
[0030] Subsequently, the learnable mask vector will be inserted into the semantic vector. and In the process of restoring the order of image patches during segmentation, this process can be represented as: in, Representative and semantic vectors The set of learnable vectors of the same dimension corresponds to image patches discarded during the mask training phase. This represents the set of subscripts for the masked image blocks. and These represent the semantic features of the student model and teacher model after being filled in, respectively, and are input into the decoders of the student model and teacher model. The output of the student model is used... Indicates. Due to Take the first one from the teacher model decoder. Output of the Layer Transformer module For each label, the decoder distillation loss is calculated, which can be expressed as: Use weights By weighting the distillation losses, the complete distillation loss can be expressed as: Based on this loss, the parameters of the student model are updated through backpropagation, thus completing one round of knowledge distillation training. Throughout the distillation process, the parameters of the teacher model remain frozen.
[0031] Execute locally After rounds of knowledge distillation, the local model The model will be uploaded to the server for model aggregation to obtain a global model. The local device downloads the global model from the server to update the local model, resulting in... This initiates a new round of federated learning. This process is repeated until the global model converges.
[0032] 3) Semantic transmission: In the second stage, the global model is preserved. encoder With the addition of a channel coding module and a semantic decoder, the complete semantic transmission system is shown in Figure 3.
[0033] The channel coding module generates semantic weights based on the current signal-to-noise ratio, representing each semantic vector. Based on their relevance to the task objective, more important semantic vectors will be allocated more communication resources. The weighted semantic vectors will then be converted into signals through power normalization. Transmission takes place. At the receiving end, the signal... Reduced to semantic features The output is obtained through semantic decoding. In Figure 3, we use This represents the distribution of data samples to be sent, where one image sample is represented as... Correspondingly, the distribution of semantic features and the distribution of task labels can be represented as follows: and ,sample The corresponding semantic features and task labels can be represented as: and .
[0034] Figure 4 illustrates the architecture of the channel coding module. The signal-to-noise ratio (SNR) is expanded into a multi-dimensional vector after passing through a fully connected neural network (MLP), and then transformed into channel weight information using the sigmoid function. Semantic features Similarly, the dimensions are expanded using an MLP, and then fused with the channel weights. The fused result is then transformed into semantic weights using a sigmoid function and combined with the original semantic features. The weighted result will be normalized to obtain the transmittable signal. .
[0035] This semantic communication system focuses on two types of tasks: image classification and image reconstruction. The loss function for semantic transmission uses the information bottleneck and can be expressed as: in, Represents the loss weight. Represents mutual information. Represents the original data and semantic features The mutual information, which is minimized, represents maximizing the data compression rate of the semantic communication system. Represents semantic features and task tags The greater the mutual information, the more accurate the semantic extraction will be, as the greater the correlation between the extracted semantic features and the task objective.
[0036] Image samples Task Tags The probability distributions can be estimated by sampling from the dataset, and are expressed as follows: and Probability distribution of semantic features If it cannot be obtained directly, a variational parameter can be introduced through reparameterization techniques. To estimate, and then obtain mutual information. The upper realm : Maximizing the upper bound minimizes mutual information loss, based on user... Local dataset The probability distribution can be estimated through sampling, resulting in the following form: in, represent Divergence. It can be estimated numerically. Divergence, as shown below: in, All are constants. This refers to the Sigmoid function. ,in This represents the variance of the channel noise.
[0037] use Mutual information represents the output of a semantic communication system. It can be estimated through a loss function specific to the task scenario. For classification tasks, this mutual information is equivalent to: in, Represents cross-entropy, at this time This represents the probability vector for the image category. For the reconstruction task, this mutual information is equivalent to: in, This represents the mean square error, at which point... This represents the original image that has been reconstructed.
[0038] 4) Model Training: This invention considers the federated learning process in heterogeneous data scenarios, using parameters... The Dirichlet distribution simulates the non-independent and identically distributed data, using This represents a semantic decoder. In image classification tasks, Composed of fully connected neural networks, it maps semantic features to predictions of image categories. In image reconstruction tasks, A similar structure to the encoder is used, employing multi-layer Transformer modules to predict pixel information of the image. Algorithm 1 demonstrates the first-stage federated learning and knowledge distillation process, while Algorithm 2 demonstrates the second-stage semantic transfer training process.
[0039] In summary, this invention designs a personalized semantic communication method based on knowledge distillation, which includes a two-stage model training process. In the first stage, a federated learning architecture is adopted to perform model pre-training on a local client. The student model, based on a local dataset, efficiently learns semantic extraction capabilities through distillation. The student model participates in the federated learning model aggregation process, which significantly reduces communication resource overhead due to its smaller scale. The student model in the first stage only contains a semantic encoder and a semantic decoder, without a channel coding module. In the second stage, the encoder from the student model in the first stage, which contains the client's personalized knowledge, is retained. Based on this, a channel coding module is added, and a suitable semantic decoder is designed according to the specific task. Semantic communication functions, including image classification and image reconstruction, are trained based on a public dataset. This scheme not only achieves personalized semantic communication functions while protecting user privacy but also effectively separates the channel coding and semantic coding processes. The semantic encoder obtained in the first stage serves as a general module, which can be supplemented with corresponding channel coding modules according to the communication environment, achieving maximum reuse. The method flow of this invention can be summarized as follows: Figure 5 As shown.
[0040] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.
Claims
1. A personalized semantic communication method based on knowledge distillation, characterized in that, Includes the following steps: S1. The local client downloads the teacher model and the local student model. The local client performs knowledge distillation and federated learning, while simultaneously inputting data into the teacher model. and student model In the middle, student model Teacher model The semantic representation is used as a label, and the semantic extraction ability is learned. After distillation, the student models are uploaded to the central server for aggregation to obtain the global model. ,go through After rounds of training, a converged global student model is obtained. Global student model It includes the user's personalized knowledge; S2, Preserve the global student model semantic encoder Furthermore, a channel coding module and a semantic decoder are added to this; let the input image be... The semantic features obtained after passing through the semantic encoder are The channel coding module is used for semantic features Channel coding is performed to obtain the encoded signal that can be transmitted. The encoded signal can be transmitted The lossy semantic signal is sent to the receiving end and received at the receiving end. The receiving end will receive the lossy semantic signal. Convert into semantic information The data is then input into the corresponding semantic decoder to complete the semantic communication task.
2. The personalized semantic communication method based on knowledge distillation according to claim 1, characterized in that: The local client is One, represented as Each client Have local dataset and local student models ,in Representing the iteration rounds of federated learning, the dataset is represented as... The number of all training samples is expressed as Teacher Model and student model Each has a pair of encoders and decoders, represented as follows: and .
3. The personalized semantic communication method based on knowledge distillation according to claim 2, characterized in that: use and The parameter weights of the global student model and the local student model are represented respectively. The aggregation process in S1 is expressed as follows: 。 4. The personalized semantic communication method based on knowledge distillation according to claim 1, characterized in that: S2 channel coding module for semantic features Channel coding is performed to obtain the encoded signal that can be transmitted. As shown below: Among them, the function Represents the channel coding module, parameters Represents the current signal-to-noise ratio. This represents the encoded signal that can be transmitted. After power normalization, the transmission process is represented as follows: in, Represents the channel gain matrix. Represents channel noise. This represents the lossy semantic signal at the receiving end.
5. A personalized semantic communication method based on knowledge distillation according to claim 2, characterized in that: In S1, the semantic extraction capabilities of the teacher model are transferred to the student model through knowledge distillation. Both the encoder and decoder of the student and teacher models are composed of Transformer modules. In the student model, the encoder and decoder respectively contain... and Each Transformer block, in the teacher model, contains the encoder and decoder respectively. and A Transformer block, and .
6. The personalized semantic communication method based on knowledge distillation according to claim 5, characterized in that: S1 specifically includes the following operations: Original image Divided into Image patches, represented as: Image mask training method is adopted, based on mask rate A fixed proportion of image patches are discarded, and the remaining image patches are used for model training, represented as... , where set This represents the set of indices of the visual image patches that were not discarded by the mask. This represents the number of image patches that are retained. The inputs are simultaneously fed into both the student model and the teacher model, resulting in semantic vectors for the student model. and teacher model semantic vector The process is represented as: The student model uses the output of the teacher model as labels for supervised learning, using a function. Representing L1 loss, the encoder distillation loss is expressed as: in, Represents a fully connected layer, used for alignment. and The dimension; The learnable mask vectors are inserted into the semantic vectors respectively. and In the process of restoring the order of image patches during segmentation, the following is represented: in, Representative and semantic vectors Learnable vectors of the same dimension correspond to image patches discarded during the mask training phase, and the set Represents the set of subscripts for the masked image blocks; and These represent the semantic features of the student model and teacher model after being filled in, respectively, and are input into the decoders of the student model and teacher model. The output of the student model is used... This indicates that the first term in the teacher model decoder is taken. Output of the Layer Transformer module For each label, the decoder distillation loss is calculated and expressed as: Use weights By weighting the distillation loss, the complete distillation loss is expressed as: Based on this loss, the parameters of the student model are updated through backpropagation, executed locally. After rounds of knowledge distillation, the local student model The model will be uploaded to the server for model aggregation to obtain a global model. The local device downloads the global model from the server to update the local student model, thus obtaining the updated model. A new round of federated learning begins, and this process is repeated until the global model converges, resulting in the global student model. .
7. The personalized semantic communication method based on knowledge distillation according to claim 1, characterized in that: The architecture of the channel coding module in S2 includes: the signal-to-noise ratio (SNR) is expanded into a multi-dimensional vector through a fully connected neural network (MLP), and then transformed into channel weight information and semantic features through a sigmoid function. After expanding the dimensions using MLP and fusing with channel weights, the fused result is transformed into semantic weights via a Sigmoid function and then combined with the original semantic features. The weighted result is then normalized to obtain the encoded, transmittable signal. .
8. The personalized semantic communication method based on knowledge distillation according to claim 2, characterized in that: The loss function for semantic transmission, employing the information bottleneck, is expressed as: in, Represents the loss weight. Represents mutual information, Represents the distribution of image samples and semantic feature distribution mutual information, Represents semantic feature distribution and task tag distribution Mutual information.
9. A personalized semantic communication method based on knowledge distillation according to claim 8, characterized in that: Image samples collected from the dataset are represented as follows: The corresponding label samples and semantic feature samples are represented as follows: and The probability distributions of image samples and task labels are estimated by sampling from the dataset and are expressed as follows: and Probability distribution of semantic features It is unknown; variation is introduced through reparameterization techniques. To estimate and obtain mutual information The upper realm : Based on user Local dataset By sampling and estimating the probability distribution, we obtain the following form: in, represent Divergence, estimated numerically Divergence, as shown below: in, For the Sigmoid function, ,in The variance represents the channel noise. Mutual Information Through loss function estimation under specific task scenarios, for classification tasks, this mutual information is equivalent to: in, Represents cross-entropy. A probability vector representing the image category; For the reconstruction task, this mutual information is equivalent to: in, Represents the mean square error. This represents the original image that has been reconstructed.