Generation method and communication method of semantic coding and decoding model based on privacy protection

Abstract

The invention discloses a semantic coding and decoding model generation method based on privacy protection, which relates to the field of semantic communication privacy protection, and comprises the following steps: acquiring semantic task requirements and shared values, and constructing a semantic coding and decoding model; generating a shared value set according to semantic coding and decoding model parameters and a local data set, and performing integer processing on the shared value set; dividing the integers of the shared value set into shared value proportions, and distributing the shared value proportions to participants; each participant inputs the shared value proportion into a deep neural network of a full connection layer, and performs iterative training of a local semantic coding model by using the held model parameter proportion and the local data proportion; parameters of the local semantic coding models are updated through a gradient noise addition iteration training process, and the trained local semantic coding models are obtained; and analyzing each local semantic coding model by using an analytic hierarchy process to obtain a corresponding weight, and carrying out weighted calculation to obtain a semantic coding and decoding model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semantic communication privacy protection, and mainly relates to a generation method and a communication method of a semantic coding and decoding model based on privacy protection. BACKGROUND

[0002] Existing semantic communication systems mainly focus on expanding application scenarios and optimizing performance, often ignoring potential privacy leakage problems that may occur in the training and inference processes of deep learning models. Common privacy threats in semantic communication systems include data leakage attacks, model overfitting, poisoning attacks, adversarial attacks, inference attacks, and model inversion attacks, which can lead to model misjudgment, information leakage, and service interruption. Therefore, semantic communication models with privacy protection are crucial.

[0003] However, in current model training practices, excessive reliance on central servers increases the risk of single point failure, making the entire training process vulnerable to central server failures or security issues. At the same time, this dependence also exacerbates the risk of data privacy leakage, as user data may be stolen or misused during upload to the central server. In addition, neglecting user data during training not only wastes data resources, but also may expose sensitive information of users to the risk of inference or leakage due to the lack of adequate privacy protection mechanisms, which poses a serious threat to user privacy and data security.

[0004] Chinese invention patent with publication number "CN115632761A" discloses a "multi-user distributed privacy protection regression method and device based on secret sharing", which specifically discloses that a data provider respectively distributes the respective privacy data , ,..., to a first server and a second server through additive arithmetic secret sharing, obtaining secret proportions of data features and data labels; the first server and the second server perform secure two-party computation based on the obtained secret proportions of data features and data labels; the first server and the second server send the respective secret proportions of model parameters and to a data user, and then the data user reconstructs the complete model parameters However, this method is limited to two-party computation and lacks flexibility, making it difficult to scale to more participants. If the computation task is changed or new functions are introduced, the entire system needs to be redesigned or extensive modifications need to be made. In addition, this method only involves two servers in the computation. If one of the servers fails or is attacked, the entire system may be affected, which can easily lead to a single point of failure. At the same time, all data needs to be transmitted between the two servers, which will result in a large communication overhead. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a method for generating and communicating a privacy-preserving semantic encoding / decoding model.

[0006] The technical solution of this application is as follows: On the one hand, this invention proposes a method for generating a privacy-preserving semantic encoding / decoding model, the method comprising: Get The semantic task requirements negotiated by the participating parties and Using the local datasets of each participant, a semantic encoding / decoding model is constructed, and the parameters of the semantic encoding / decoding model are initialized according to the semantic task requirements. Based on the semantic encoding / decoding model parameters and the local datasets, a shared value set is generated, and the shared value set is integerized. The integerized shared value set is then divided into... The proportion of shared value is distributed to One participating party; Each participant inputs the shared value weight into a fully connected deep neural network, and uses its held model parameter weights to iteratively train its local semantic coding model. During the iterative training process, the parameters of the local semantic coding model are updated to obtain a trained model. A local semantic encoding model; Using the Analytic Hierarchy Process (AHP) The local semantic coding model is analyzed to obtain the corresponding weights. The weighted calculation is performed based on the local semantic coding model and the corresponding weights to obtain the final semantic coding and decoding model.

[0007] Preferably, a semantic encoding / decoding model is constructed, specifically as follows: based on A deep neural network with fully connected layers constructs a semantic encoder-decoder model, the parameters of which include semantic encoder parameters. Semantic decoder parameters and pragmatic function parameters According to the requirements of the semantic task, The participating parties jointly negotiated and determined the initial set of semantic encoding and decoding model parameters. ,in, Indicates the first The index value of each participant; Indicates the first The first participating party In the weight matrix of layer 1 Line number The value of the column; Indicates the weighted row index; Indicates the index of the weighted column; Indicates the first The first participating party Layer tendency parameters; This represents the initial set of parameters for the semantic encoding / decoding training model.

[0008] Preferably, a shared value set is generated based on the semantic encoding / decoding model parameters and the local dataset, and the shared value set is then integerized, specifically as follows: The The local datasets of each participant are ,in, Indicates the first Local datasets of each participant; Based on semantic encoding and decoding model parameters and Merge the local datasets to obtain a shared value set. ,in, This represents the union operation; Extract the small values ​​from the shared value set, and convert them to integers, as shown by the formula: , ; ; ; In the formula, Represents the integer after integerization; Represents the first value in the shared value set. A small number; Indicates the scaling factor; Indicates the preset number of decimal places; Indicates the number of decimal values ​​in the shared value set; Indicates the first Index values ​​of data in a local dataset; Represents the set of shared values ​​after integerization; This represents the number of elements in the integerized set of shared values. Based on a preset set of zero weights, the integerized shared value set is encrypted using modular arithmetic and divided into... The proportions are shared, expressed by the formula: ; ; In the formula, Indicates the first The shared value held by each participant; Indicates the first Each participant will share the value. To allocate to oneself; Indicates the first Each participant will share the value. The zero weight allocated to oneself; Indicates the first Each participant will share the value. Assigned to the Zero weight for each participating party; Indicates the first Each participant will share the value. Assigned to the The proportion of shared value among each participant; The modulus representing the range of calculations; This indicates a modulo operation.

[0009] Preferably, each participant inputs the shared value weight into a fully connected deep neural network, and encrypts the input shared value weight based on a trusted tuple provided by a trusted third party, generating an encrypted weight, expressed by the formula: ; ; ; In the formula, Indicates the first All that each participant receives A set of shared value proportions; Indicates the first The set of shared value proportions of each participant The proportion of shared values; and Indicates the first The proportion of credible binary groups among the participants; Indicates the first The set of shared value proportions of each participant The first in The encryption ratio of each shared value; Indicates the first All that each participant receives A set of encrypted weights for shared values, wherein the set of encrypted weights Including the The encrypted proportion of different row and column elements in the weight matrix of the participant, the encrypted proportion of the bias parameter, and the encrypted proportion of the real value of the shared value; and represents the trusted binary tuple of the th participant; The participant performs forward propagation calculation using the data encryption proportion and the model parameter proportion, and the formula is expressed as: ; In the formula, represents the predicted value of the th participant's shared value of the th layer; represents the predicted value of the th participant's shared value of the th layer; represents the weight matrix encryption proportion of the th participant's weight matrix of the th layer; represents the bias parameter encryption proportion of the th participant; represents the real value of the shared value of the th participant; represents the real value of the shared value of the th participant;

[0010] represents the activation function of the hidden layer. Preferably, in the step of updating the parameters of the local semantic coding model, the loss function is used for back propagation, and the gradient is updated, and the formula is expressed as: In the formula, represents the predicted value of the th participant's shared value; The th participant calculates the weight matrix gradient and the bias parameter gradient of each layer, wherein the bias parameter gradient of the th layer full connection layer is calculated through error information, and the formula is expressed as: ; In the formula, represents the predicted value of the th participant's shared value of the th layer output layer; represents the real value of the th participant's shared value of the th layer full connection layer; represents the bias parameter gradient of the th participant's full connection layer of the th layer; Indicates the first Error information of fully connected layers; Indicates the number of output layers; Indicates the transpose operation; The gradient of the weight matrix is ​​expressed by the formula: ; In the formula, Indicates the first The first participating party Gradient of the layer weight matrix.

[0011] Preferably, the parameters of the updated local semantic coding model are obtained to obtain the trained model. A local semantic encoding model, specifically: The gradient of the local semantic coding model parameters is clipped by setting a maximum gradient norm, as expressed by the formula: ; ; ; In the formula, The gradient vector is the set of gradients representing the gradients of the weight matrix and the gradients of the bias parameter. Indicates the first The index values ​​of the gradient vectors; Indicates the first The components of the gradient vector; Represents the gradient norm; Indicates the total number of model parameters; This represents the preset maximum gradient norm; Indicates the gradient after clipping; Based on the additive noise mechanism of differential privacy, it can be expressed by the following formula: ; In the formula, Indicates the standard deviation of noise; Indicates the preset privacy sensitivity; This indicates the preset probability of privacy leakage; Indicates the preset batch size; This indicates the preset privacy parameters; Calculate the gradient after adding noise, and update the gradient of the local semantic coding model parameters, expressed by the formula: ; ; ; ; ; ; wherein, denotes the model parameter gradient after adding noise; denotes the normal distribution of the noise standard deviation; denotes the noise standard deviation; denotes the updated semantic encoder parameter; denotes the updated semantic decoder parameter; denotes the updated pragmatic function parameter; denotes the new weight matrix of the layer of the th participant; denotes the new bias parameter of the layer of the th participant; denotes the learning rate.

[0012] Preferably, the local semantic encoding models of the participants are analyzed by using the analytic hierarchy process to obtain the corresponding weights, and the importance of each two local semantic encoding models is scored according to an expert scoring table to obtain a score matrix , which is specifically expressed as: , , ; wherein, denotes the score value of the importance of each two local semantic encoding models in the th row and the th column, denotes the score value of the importance of each two local semantic encoding models in the th row and the th column, wherein and are reciprocal of each other; denotes the row index value of the th row score matrix; denotes the column index value of the th column score matrix; solves the product of each row of the matrix to the power to obtain a weight vector of dimensions, which is expressed by the formula as: ; is normalized to obtain the corresponding weight of the local semantic encoding model, which is expressed by the formula as: ; wherein, Indicates the first The weights of each local semantic coding model; Indicates the first The weight vector of a local semantic coding model; The final semantic encoding / decoding model is obtained by weighting the local semantic encoding model and its corresponding weights, as expressed by the formula: ; In the formula, This represents the final semantic encoding / decoding model; Indicates the first A local semantic encoding model.

[0013] On the other hand, the present invention also proposes a semantic communication method, which uses a semantic encoding and decoding model generated by the privacy-preserving semantic encoding and decoding model generation method as described in any embodiment of the present invention. The semantic communication method includes the following steps: Communicating in response to a semantic communication request from any participant, the semantic communication request carrying source information of the participant; Based on the semantic encoding and decoding model, semantic extraction is performed on the source information to obtain semantically extracted information; Send semantic extraction information to the participants.

[0014] In another aspect, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for generating a privacy-preserving semantic encoding and decoding model as described in any embodiment of the present invention.

[0015] In another aspect, the present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for generating a privacy-preserving semantic encoding and decoding model as described in any embodiment of the present invention.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1) This invention provides a method for generating a privacy-preserving semantic encoding and decoding model. It encrypts the local dataset using an arithmetic secret sharing algorithm to generate shared values, and distributes these shared values ​​and semantic encoding and decoding model parameters to various participants in multiple weighted proportions. This enhances the protection of privacy data, reduces the risk of data leakage, and ensures the privacy and security of data during transmission. 2) This invention provides a method for generating a privacy-preserving semantic encoding and decoding model. By pruning the gradient by presetting the maximum gradient norm, the gradient explosion problem is avoided, the training stability is improved, the training efficiency of the model is improved, and the accuracy and generalization ability of the model are enhanced. 3) This invention provides a method for generating a privacy-preserving semantic encoding and decoding model. It uses the hierarchical analysis method to analyze each local semantic encoding model, obtains the corresponding weights, improves the adaptability and flexibility of the model, enhances the accuracy and robustness of the model, and strengthens the personalization and customization capabilities of the model. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0019] This invention provides the following technical solution: a method for generating a privacy-preserving semantic encoding and decoding model.

[0020] Example 1 See details Figure 1 This embodiment provides a method for generating a privacy-preserving semantic encoding and decoding model, the specific steps of which include: S1, Obtain The semantic task requirements negotiated by the participating parties and Using the local datasets of each participant, construct a semantic encoding and decoding model, and initialize the parameters of the semantic encoding and decoding model according to the requirements of the semantic task; S11, based on A deep neural network with fully connected layers constructs a semantic encoder-decoder model, the parameters of which include semantic encoder parameters. Semantic decoder parameters and pragmatic function parameters According to the requirements of the semantic task, The participating parties jointly negotiated and determined the initial set of semantic encoding and decoding model parameters. ,in, Indicates the first The index value of each participant; Indicates the first The first participating party In the weight matrix of layer 1 Line number The value of the column; Indicates the weighted row index; Indicates the index of the weighted column; Indicates the first The first participating party Layer tendency parameters; This represents the initial set of parameters for the semantic encoding / decoding training model. S2. Generate a shared value set based on the semantic encoding / decoding model parameters and the local dataset, and perform integerization on the shared value set; The The local datasets of each participant are ,in, Indicates the first Local datasets of each participant; Based on semantic encoding and decoding model parameters and Merge the local datasets to obtain a shared value set. ,in, This represents the union operation; Extract the small values ​​from the shared value set, and convert them to integers, as shown by the formula: , ; ; ; In the formula, Represents the integer after integerization; Represents the first value in the shared value set. A small number; Indicates the scaling factor; Indicates the preset number of decimal places; Indicates the number of decimal values ​​in the shared value set; Indicates the first Index values ​​of data in a local dataset; Represents the set of shared values ​​after integerization; This represents the number of data items in the integerized shared value set; Based on a preset set of zero weights, the integerized shared value set is encrypted using modular arithmetic and divided into... The proportions are shared, expressed by the formula: ; ; In the formula, Indicates the first The shared value held by each participant; Indicates the first Each participant will share the value. The zero weight allocated to oneself; Indicates the first Each participant will share the value. zero proportion allocated to the first participating party; denotes the shared value allocated to the first participating party; denotes the shared value proportion allocated to the first participating party; denotes the modulus of the calculation range; denotes the modulo operation; S3, divide the integerized shared value set into shared value proportions and distribute them to participating parties; S4, each participating party inputs the shared value proportion into a deep neural network of a fully connected layer, and performs local semantic encoding model training using the held model parameter proportion; The deep neural network mainly consists of fully connected layers, and improves the efficiency of subsequent multi-party calculation and privacy protection by intelligently encoding the input data. layers, each layer has nodes, wherein ; the semantic encoder and decoder do not contain activation functions, and each fully connected layer of the pragmatic task function is added with an activation function ; S41, the training process includes forward propagation calculation and back propagation calculation; S411, based on the forward propagation calculation, a predicted value of the shared value is obtained; Each participating party inputs the shared value proportion into a deep neural network of a fully connected layer, encrypts the input shared value proportion based on the trusted binary tuple provided by the trusted third party, and generates an encrypted proportion, which is expressed by the formula: ; ; ; In the formula, denotes the set of all shared value proportions obtained by the first participating party; denotes the shared value proportion in the set of shared value proportions of the first participating party; and denote the proportion of the trusted binary tuple of the first participating party; denotes the shared value proportion in the set of shared value proportions of the first participating party The encryption ratio of each shared value; Indicates the first All that each participant receives A set of encrypted weights for shared values, wherein the set of encrypted weights Including the The encryption weight of different row and column elements in the weight matrix of each participant, the encryption weight of the bias parameter, and the encryption weight of the true value of the shared value; and Indicates the first A trusted binary group of each participant; The participants perform forward propagation calculations using the weighting of data encryption and model parameters, expressed by the formula: ; In the formula, Indicates the first The first participating party Predicted values ​​of layer shared values; Indicates the first The first participating party The weight matrix of a layer is formed by the encryption weight of elements in different rows and columns; Indicates the first The first participating party The encryption weight of the layer's tendency parameters; Indicates the first The encrypted proportion of the true value shared by each participant; Indicates the first The true value shared by each participant; This represents the activation function of the hidden layer; During the generation of encrypted weights, the privacy of input data is guaranteed through encryption. All operations are performed only on encrypted data to ensure that each participant does not access the data content of other participants. Each participant only performs local calculations, but encrypted data exchange is required between different participants. Each participant needs to exchange training results. S42. In the step of updating the parameters of the local semantic coding model, based on the loss function... Perform backpropagation and update the gradient, expressed by the formula: ; In the formula, Indicates the first The predicted value shared by each participant; No. Each participant calculates the gradient of the weight matrix and the gradient of the bias parameter for each output layer, wherein the gradient of the first layer is obtained through error information. The gradient of the layer's tendency parameter, specifically error information, is expressed by the formula: ; In the formula, Indicates the first The first participating party The layer outputs the predicted value of the shared layer value; Indicates the first The first participating party The actual value of the shared value in the fully connected layer; Indicates the first The first participating party The gradient of the tendency parameter of the fully connected layer; Indicates the first Error information of fully connected layers; Indicates the number of output layers; Indicates the transpose operation; The gradient of the weight matrix is ​​expressed by the formula: ; In the formula, Indicates the first The first participating party Gradient of the layer's weight matrix; In another implementation, the local semantic coding model is updated using a composite loss function based on semantic task requirements. For example, if the semantic task requirements include image classification, then the composite loss function... , constructed as: ; In the formula, Represents the image restoration loss function; This represents the image classification loss function; The hyperparameters representing the recovery loss function; The hyperparameters of the classification loss function are represented. Hyperparameters representing the tradeoff between observational and pragmatic information; The recovery loss function and classification loss function are expressed by the following formulas: ; ; In the formula, Indicates the total number of data samples; This represents the total number of categories in the classification task; Indicates the first The first The true classification result of class shared values; Indicates the first The first Predicted classification results for class shared values; denotes a logarithm function; wherein the value of is calculated by iteration, and is expressed as: wherein, denotes the predicted classification result of the shared value after the th iteration; until the convergence condition is met wherein denotes a preset convergence precision; the cross-entropy loss function provides the required logarithm calculation result; The composite function is regularized, and is expressed as: wherein, denotes a hyperparameter of the regularization of the composite loss function; The local semantic encoding model parameters are updated by the composite loss function in the subsequent steps, which will not be described here; S5, the gradient update is specifically clipping the gradient of the local semantic encoding model parameters with a preset maximum gradient norm, and is expressed as: wherein, denotes a gradient set, i.e., a gradient vector, of the gradient of the weight matrix and the gradient of the inclination parameter; denotes the index value of the th gradient vector; denotes the component of the th gradient vector; denotes the gradient norm; denotes the total number of model parameters; denotes a preset maximum gradient norm; denotes the clipped gradient; wherein, it is ensured that the semantic information of the data is not lost in the clipping operation, the gradient clipping operation is performed in an encrypted state, and the clipped gradient is still encrypted; S6, the gradient is added with noise based on the additive noise mechanism of differential privacy, to generate a noise gradient, which is expressed as: wherein, denotes a noise standard deviation; denotes a preset privacy sensitivity; denotes a preset privacy leakage probability; ​​​​​​​​​​represents a preset batch size; represents a preset privacy parameter; The gradient after adding noise is calculated, and the local semantic encoding model parameter gradient is updated, which is expressed by the formula: ; ; ; ; ; ; In the formula, represents the model parameter gradient after adding noise; represents a normal distribution of noise standard deviation; represents the noise standard deviation; represents the updated semantic encoder parameter; represents the updated semantic decoder parameter; represents the updated semantic function parameter; represents the new weight matrix of the layer of the th participant; represents the new bias parameter of the layer of the th participant; represents the learning rate; S7, updating the gradient after adding noise to the local semantic encoding model parameter , the parameter of the local semantic encoding model is updated in the iterative training process, and local semantic encoding models trained are obtained; S8, using the analytic hierarchy process to analyze local semantic encoding models to obtain corresponding weights, and performing weighted calculation according to the local semantic encoding model and the corresponding weights to obtain a semantic encoding and decoding model; According to the expert scoring table, as shown in Table 1, the importance of each two local semantic encoding models is scored to obtain a score matrix , which is specifically expressed as: , , ; In the formula, represents the score value of the importance of each two local semantic encoding models in the th row and the th column, represents the score value of the importance of each two local semantic encoding models in the th row and the The scoring value of the importance of each local semantic coding model column, wherein and are reciprocal to each other; denotes the row index value of the row score matrix; denotes the column index value of the column score matrix; Table 1 Expert Scoring Table

[0021] The power of the product of each row of the matrix is solved to obtain a weight vector of dimension, which is expressed by the formula as follows: ; Normalization processing is performed to obtain the corresponding weight of the local semantic coding model, which is expressed by the formula as follows: ; In the formula, denotes the weight of the th local semantic coding model; denotes the weight vector of the th local semantic coding model; According to the local semantic coding model and the corresponding weight, a weighted calculation is performed to obtain a final semantic coding and decoding model, which is expressed by the formula as follows: ; In the formula, denotes the final semantic coding and decoding model; denotes the th local semantic coding model. Embodiment 2

[0022] The embodiment provides a semantic communication method, which adopts the semantic coding and decoding model generated by the generation method of the semantic coding and decoding model based on privacy protection in any embodiment, and the semantic communication method comprises the following steps: communicating in response to a semantic communication request of any participant, wherein the semantic communication request carries source information of the participant; performing semantic extraction on the source information based on the semantic coding and decoding model to obtain semantic extraction information; sending the semantic extraction information to the participant.

[0023] Embodiment 3 ​The embodiment provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the generation method of the privacy protection based semantic coding and decoding model according to any embodiment of the present application when executing the program.

[0024] Embodiment 4 The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the generation method of the privacy protection based semantic coding and decoding model according to any embodiment of the present application.

[0025] It is worth noting that the electronic device and the computer readable storage medium according to the present application are based on the same principle as the method according to the embodiment 1, and will not be described here.

[0026] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation according to the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for generating a privacy-preserving semantic encoding / decoding model, characterized in that, The method includes: Get Semantic task requirements negotiated by the participating parties and Using the local datasets of each participant, a semantic encoding / decoding model is constructed, and the parameters of the semantic encoding / decoding model are initialized according to the semantic task requirements. Based on the semantic encoding / decoding model parameters and the local datasets, a shared value set is generated, and the shared value set is integerized. The integerized shared value set is then divided into... The proportion of shared value is distributed to One participating party; Each participant inputs the shared value weight into a fully connected deep neural network, and uses its held model parameter weights to iteratively train its local semantic coding model. During the iterative training process, the parameters of the local semantic coding model are updated to obtain a trained model. A local semantic encoding model; Using the Analytic Hierarchy Process (AHP) The local semantic coding model is analyzed to obtain the corresponding weights. The weighted calculation is performed based on the local semantic coding model and the corresponding weights to obtain the final semantic coding and decoding model.

2. The method for generating a privacy-preserving semantic encoding / decoding model according to claim 1, characterized in that, Constructing a semantic encoding / decoding model, specifically: based on A deep neural network with fully connected layers constructs a semantic encoder-decoder model, the parameters of which include semantic encoder parameters. Semantic decoder parameters and pragmatic function parameters According to the requirements of the semantic task, The participating parties jointly negotiated and determined the initial set of semantic encoding and decoding model parameters. ,in, Indicates the first The index value of each participant; Indicates the first The first participating party In the weight matrix of layer 1 Line 1 The value of the column; Indicates the weighted row index; Indicates the index of the weighted column; Indicates the first The first participating party Layer tendency parameters; This represents the initial set of parameters for the semantic encoding / decoding training model.

3. The method for generating a privacy-preserving semantic encoding / decoding model according to claim 2, characterized in that, Based on the semantic encoding / decoding model parameters and the local dataset, a shared value set is generated, and the shared value set is then integerized. Specifically: The The local datasets of each participant are ,in, Indicates the first Local datasets of each participant; Based on semantic encoding and decoding model parameters and Merge the local datasets to obtain a shared value set. ,in, This represents the union operation; Extract the small values ​​from the shared value set, and convert them to integers, as shown by the formula: ， ； ； ； In the formula, Represents the integer after integerization; Represents the first value in the shared value set. A small number; Indicates the scaling factor; Indicates the preset number of decimal places; Indicates the number of decimal values ​​in the shared value set; Indicates the first Index values ​​of data in a local dataset; Represents the set of shared values ​​after integerization; This represents the number of elements in the integerized set of shared values. Based on a preset set of zero weights, the integerized shared value set is encrypted using modular arithmetic and divided into... The proportions are shared, expressed by the formula: ； ； In the formula, Indicates the first The shared value held by each participant; Indicates the first Each participant will share the value. To allocate to oneself; Indicates the first Each participant will share the value. The zero weight allocated to oneself; Indicates the first Each participant will share the value. Assigned to the Zero weight for each participating party; Indicates the first Each participant will share the value. Assigned to the The proportion of shared value among each participant; The modulus representing the range of calculations; This indicates a modulo operation.

4. The method for generating a privacy-preserving semantic encoding / decoding model according to claim 3, characterized in that, Each participant inputs the shared value weight into a fully connected deep neural network. Based on a trusted tuple provided by a trusted third party, the input shared value weight is encrypted to generate an encrypted weight, expressed by the formula: ； ； ； In the formula, Indicates the first All that each participant receives A set of shared value proportions; Indicates the first The set of shared value proportions of each participant The proportion of shared values; and Indicates the first The proportion of credible binary groups among the participants; Indicates the first The set of shared value proportions of each participant The first in The encryption ratio of each shared value; Indicates the first All that each participant receives A set of encrypted weights for shared values, wherein the set of encrypted weights Including the The encryption weight of different row and column elements in the weight matrix of each participant, the encryption weight of the bias parameter, and the encryption weight of the true value of the shared value; and Indicates the first A trusted binary group of each participant; The participants perform forward propagation calculations using the weighting of data encryption and model parameters, expressed by the formula: ； In the formula, Indicates the first The first participating party Predicted values ​​of layer shared values; Indicates the first The first participating party The weight matrix of a layer is formed by the encryption weight of elements in different rows and columns; Indicates the first The first participating party The encryption weight of the layer's tendency parameters; Indicates the first The encrypted proportion of the true value shared by each participant; Indicates the first The true value shared by each participant; This represents the activation function of the hidden layer.

5. The method for generating a privacy-preserving semantic encoding / decoding model according to claim 4, characterized in that, In the step of updating the parameters of the local semantic coding model, based on the loss function Perform backpropagation and update the gradient, expressed by the formula: ； In the formula, Indicates the first The predicted value shared by each participant; No. Each participant calculates the gradient of the weight matrix and the gradient of the bias parameter for each layer, wherein the gradient of the first layer is calculated using error information. The gradient of the tendency parameter of a fully connected layer is expressed by the formula: ； In the formula, Indicates the first The first participating party The layer outputs the predicted value of the shared layer value; Indicates the first The first participating party The actual value of the shared value in the fully connected layer; Indicates the first The first participating party The gradient of the tendency parameter of the fully connected layer; Indicates the first Error information of fully connected layers; Indicates the number of output layers; Indicates the transpose operation; The gradient of the weight matrix is ​​expressed by the formula: ； In the formula, Indicates the first The first participating party Gradient of the layer weight matrix.

6. The method for generating a privacy-preserving semantic encoding / decoding model according to claim 5, characterized in that, The parameters of the updated local semantic coding model are obtained to obtain the trained model. A local semantic encoding model, specifically: The gradient of the local semantic coding model parameters is clipped by setting a maximum gradient norm, as expressed by the formula: ； ； ； In the formula, The gradient vector is the set of gradients representing the gradients of the weight matrix and the gradients of the bias parameter. Indicates the first The index values ​​of the gradient vectors; Indicates the first The components of the gradient vector; Represents the gradient norm; Indicates the total number of model parameters; This represents the preset maximum gradient norm; Indicates the gradient after clipping; Based on the additive noise mechanism of differential privacy, it can be expressed by the following formula: ； In the formula, Indicates the standard deviation of noise; Indicates the preset privacy sensitivity; This indicates the preset probability of privacy leakage; Indicates the preset batch size; This indicates the preset privacy parameters; Calculate the gradient after adding noise, and update the gradient of the local semantic coding model parameters, expressed by the formula: ； ； ； ； ； ； In the formula, This represents the gradient of the model parameters after adding noise; The normal distribution representing the standard deviation of noise; Indicates the standard deviation of noise; This represents the updated semantic encoder parameters; This represents the updated semantic decoder parameters; This represents the updated pragmatic function parameters; Indicates the first The first participating party The new weight matrix of the layer; Indicates the first The first participating party New tendency parameters for the layer; This represents the learning rate.

7. The method for generating a privacy-preserving semantic encoding / decoding model according to claim 1, characterized in that, Using the Analytic Hierarchy Process (AHP) The local semantic coding models of each participant are analyzed to obtain corresponding weights. The importance of each pair of local semantic coding models is scored according to an expert scoring table, resulting in a score matrix. Specifically, it is expressed as: ， ， ； In the formula, Indicates the first Line 1 The column lists the importance scores for each pair of local semantic coding models. Indicates the first Line 1 The column lists the scores for the importance of each pair of local semantic coding models, where... and They are reciprocals of each other; Indicates the first The row index value of the row score matrix; Indicates the first Column index values ​​of the column score matrix; Solve for the product of each row of the matrix The power of 1 yields a dimensional weight vector This can be expressed as a formula: ； After normalization, the corresponding weights of the local semantic coding model are obtained, expressed by the formula: ； In the formula, Indicates the first The weights of each local semantic coding model; Indicates the first The weight vector of a local semantic coding model; The final semantic encoding / decoding model is obtained by weighting the local semantic encoding model and its corresponding weights, as expressed by the formula: ； In the formula, This represents the final semantic encoding / decoding model; Indicates the first A local semantic encoding model.

8. A semantic communication method, characterized in that, A semantic codec model generated using the method for generating a privacy-preserving semantic codec model as described in any one of claims 1 to 7, wherein the semantic communication method includes the following steps: Communicating in response to a semantic communication request from any participant, the semantic communication request carrying source information of the participant; Based on the semantic encoding and decoding model, semantic extraction is performed on the source information to obtain semantically extracted information; Send semantic extraction information to the participants.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for generating a privacy-preserving semantic codec model as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method for generating a privacy-preserving semantic codec model as described in any one of claims 1 to 7.

Images