A multi-department government affair collaboration method, device, medium and product

Abstract

The application provides a multi-department government affair collaboration method, device, medium and product, and relates to the technical field of data security.The method comprises the following steps: dimension reduction is performed on an original data matrix input in a different department government affair collaboration process to generate a feature vector of the original data matrix of each department; a KL divergence value is calculated to determine the privacy protection degree of each feature vector; an n-dimensional feature vector matrix is obtained by splicing the feature vectors of each department; the feature vector matrix is combined with a rank-deficient matrix with a preset rank of n-1 to obtain a joint feature matrix; and a loss function for balancing the privacy protection degree and information loss of the original data is constructed to determine the feature vector and joint feature matrix that minimize the loss function value.The joint feature matrix has no unique solution in the application, thereby solving the problem that the model parameters are easily inversely deduced to reconstruct the original data, and the constructed loss function solves the problem that the privacy protection and information loss of the original data are difficult to balance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data security, in particular to a multi-department government affair collaboration method, device, medium and product. BACKGROUND

[0002] With the continuous advancement of the process of digital government affairs, cross-domain data collaboration between different departments becomes increasingly frequent. In emergency tasks and other scenarios, timely and accurate data collaboration is crucial, as it directly relates to the effectiveness and timeliness of response measures. However, current data collaboration faces many challenges. On the one hand, data often involves sensitive issues such as citizen privacy and institutional authority, and manual approval or desensitization processing is often used. Manual approval processes are tedious and require a lot of time and labor costs, resulting in extremely low collaboration efficiency, making it difficult to meet the urgent requirements of emergency tasks; on the other hand, while desensitization processing protects privacy to some extent, it causes serious loss of information, greatly compromising the integrity and accuracy of data, and thus affecting the precision of decision-making, which may lead to biased or erroneous decisions.

[0003] To solve the problem of data privacy and collaboration, the traditional federated learning method emerges as the times require. This method ensures data security by implementing "data does not leave the domain", that is, the original data remains in each department's local and is not directly transmitted to the outside, and only model parameters, calculation results and other ways are shared for collaboration, so that each participant can train the model while protecting local data privacy. However, this method is not perfect and still has the risk of model parameter leakage, and attackers may be able to infer relevant information about the original data by analyzing the model parameters, posing a potential threat to data security. Secure multi-party computation performs well in terms of security and can ensure that participants complete the calculation task without leaking their own data, but its computational complexity increases exponentially with the number of participants. In the context of large models, the amount of data to be processed is huge and complex, making it difficult for secure multi-party computation to achieve real-time response and unable to meet the demand for fast data processing in actual applications. SUMMARY

[0004] Embodiments of the present application provide a multi-department government affair collaboration method, device, medium and product, aiming to solve the problem that the model parameters of existing government data collaboration are easily reconstructed by reverse deduction of the original data, and the privacy protection and information loss of the original data are difficult to balance.

[0005] To achieve the above-mentioned purpose, in a first aspect, the present application provides a multi-department government affair collaboration method, comprising the following steps:

[0006] obtaining the original data matrix input by a plurality of departments participating in government affair collaboration;

[0007] The original data matrix is dimensionally reduced by an encoder to generate a feature vector of the original data matrix of each department, wherein:

[0008] ,

[0009] In the formula, Z is the feature vector of the department; X is the original data matrix; W e is the weight matrix of the encoder; b e is the bias vector; RELU is the activation function; represents a mapping function from the original data matrix X to the virtual feature vector Z;

[0010] The privacy protection degree of the feature vector is determined by calculating a KL divergence value; the KL divergence value is calculated by the following formula:

[0011] ,

[0012] ,

[0013] In the formula, represents the KL divergence function between the encoding distribution and the reference distribution ; μ is the mean vector of the feature vector Z ; Σ is the covariance matrix of the feature vector Z ; is a preset KL divergence threshold value; k is the dimension of the original data matrix; represents the encoding distribution; represents the reference distribution; represents a normal distribution;

[0014] An n-dimensional feature vector matrix is obtained by splicing the feature vectors of each department, and the feature vector matrix is combined with a rank-deficient matrix with a preset rank of n-1 to obtain a joint feature matrix, wherein:

[0015] ,

[0016] In the formula, F represents the multi-department joint feature matrix, respectively represent the virtual feature vectors of the n government departments, represents a row splicing operation, that is, the joint matrix formed by horizontally splicing the feature matrices of different departments; M is a rank-deficient matrix, wherein ;

[0017] By constructing a loss function to balance the degree of privacy protection of the original data and information loss, the feature vector and joint feature matrix that minimize the value of the loss function are determined. The loss function is calculated using the following formula:

[0018] ,

[0019] In the formula, L The loss function; Y Label the target decision; W d The decoder weight matrix is ​​the matrix corresponding to the joint feature matrix; This indicates information loss; λ γ is the privacy protection strength coefficient; γ is the feature sparsity coefficient; Represents the encoder weight matrix W e The L1 norm;

[0020] The joint feature matrix is ​​sent to multiple departments participating in government collaboration, so that these departments can make collaborative decisions based on the joint feature matrix.

[0021] Furthermore, the original data matrix also identifies key features through a two-layer attention mechanism, including:

[0022] The local attention weights are calculated using the following formula:

[0023] ,

[0024] In the formula, For the original data matrix, the first... One feature; x j For the original data matrix, the first... j One characteristic, d The total number of features in the original data matrix; Represents the original data matrix. Local attention weights for each feature; W a This is the attention weight matrix;

[0025] The global attention weights are calculated using the following formula:

[0026] ,

[0027] In the formula, Represents the original data matrix. Global attention weights for each feature; Represents the original data matrix. Features Cosine similarity to the target decision label Y; This represents the maximum cosine similarity between all features of the original data matrix and the target decision label;

[0028] A comprehensive weight is calculated using the local attention weights and the total attention weights. Features with a comprehensive weight greater than or equal to a preset comprehensive weight threshold are identified as key features of the original data matrix. The comprehensive weight is calculated using the following formula:

[0029] ,

[0030] In the formula, Represents the original data matrix. The combined weight of each feature.

[0031] Furthermore, before dimensionality reduction via the encoder, the original data matrix is ​​adjusted in dimensionality based on feedback from changes in information loss and privacy protection levels, calculated using the following formula:

[0032] ,

[0033] In the formula, The dimensions of the original data matrix after adjustment; The dimensions of the original data matrix before adjustment; Indicates the change in information loss; This represents the change in the level of privacy protection. This represents the learning rate.

[0034] Furthermore, the joint feature matrix is ​​also verified for invertibility using the following formula:

[0035] ,

[0036] In the formula, Represents the matrix condition number; The Jacobian matrix representing the joint characteristic matrix; This represents the largest singular value of the Jacobian matrix; Represents the smallest singular value of the Jacobian matrix;

[0037] when When the value is greater than the preset value, it indicates that the verification result of the invertibility of the joint feature matrix meets the standard.

[0038] Furthermore, the parameters, including the encoder weight matrix, decoder weight matrix, and bias vector, are adjusted using gradient descent and adjusted according to the following formula:

[0039] ;

[0040] In the formula, is an adjusted parameter; is a parameter before adjustment; is a learning rate; is an effect gain; represents a gradient of a loss function; represents a gain coefficient.

[0041] Further, the collaborative decision is determined according to a joint decision probability distribution of the joint feature matrix, and the collaborative decision is calculated by the following formula:

[0042]

[0043] In the formula, S is a probability distribution after the joint feature matrix is mapped by a decision weight matrix; W s is a decision weight matrix;

[0044] The decision with the maximum probability is obtained through the probability distribution as the best decision.

[0045] Further, the best decision also comprehensively scores the decision effect according to a degree of privacy protection, and the comprehensive score is calculated by the following formula:

[0046]

[0047] In the formula, represents a decision effect score, represents a prediction accuracy; is a time consumption of a traditional decision; is a time consumption of the best decision; is a KL divergence value;

[0048] When the decision effect score is greater than a preset score threshold, it indicates that the best decision effect is better than the traditional decision.

[0049] In a second aspect, the present application provides an electronic device, comprising a memory and a processor, the memory stores at least one program, and the at least one program is executed by the processor to realize the multi-department government affair collaboration method as described above.

[0050] In a third aspect, the present application provides a computer readable storage medium, the storage medium stores at least one program, and the at least one program is executed by the processor to realize the multi-department government affair collaboration method as described above.

[0051] In a fourth aspect, the present application provides a computer program product, comprising a computer program, and the computer program is executed by the processor to realize the steps of the multi-department government affair collaboration method as described above.

[0052] ​​The above technical solution has the following technical effects:

[0053] By obtaining the original data matrix input by multiple departments participating in government affair collaboration, dimension reduction is performed to generate a feature vector of the original data matrix of each department; the KL divergence value is calculated to determine the privacy protection degree of each feature vector; the feature vectors of each department are spliced to obtain an n-dimensional feature vector matrix; the feature vector matrix is combined with a rank-deficient matrix with a preset rank of n-1 to obtain a joint feature matrix; and a loss function for balancing the privacy protection degree and information loss of the original data is constructed to determine the feature vector and joint feature matrix that minimize the loss function value. In the present application, the joint feature matrix has no unique solution, solving the problem that the model parameters are easily inversely deduced to reconstruct the original data, and the constructed loss function solves the problem that the privacy protection and information loss of the original data are difficult to balance.

[0054] In further embodiments, the present application further identifies key features of the original data matrix through a double-layer attention mechanism, filters redundant features such as identifiers, adjusts the dimension of the original data matrix by combining the feedback of the information loss and the privacy protection degree variation, and realizes the optimal balance between privacy protection and information integrity.

[0055] In further embodiments, the probability distribution is used to obtain the decision with the maximum probability as the best decision, determine the best decision of the joint of different departments, and comprehensively score the decision effect according to the joint privacy protection degree to judge the effective degree of the best decision relative to the traditional decision, thereby improving the quality of government affair collaboration. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 FIG. 1 is a flowchart of a multi-department government affair collaboration method according to an embodiment of the present application;

[0057] Figure 2 FIG. 4 is a structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0058] To further illustrate the embodiments, the present application provides accompanying drawings. These drawings are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. Those of ordinary skill in the art should be able to understand other possible implementations and advantages of the present application by referring to these contents. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0059] The present application will be further described in conjunction with the accompanying drawings and specific embodiments.

[0060] Embodiment One:

[0061] Figure 1A flowchart of a multi-department government affair collaboration method according to an embodiment of the present application is shown in FIG. 1, which includes the following steps: Figure 1

[0062] Obtaining an original data matrix input by a plurality of departments participating in government affair collaboration;

[0063] In a specific implementation, the government affair collaboration according to the embodiment takes a flood emergency material intelligent scheduling scenario as an example, and needs to quickly coordinate real-time disaster data such as population distribution and house damage rate of a township of the civil affairs department, and inventory data such as tents, lifeboats, and medicines of b material reserve points of the emergency management department, for example,

[0064] Civil affairs department data matrix , where n1 is the sample number and d1 is the feature number;

[0065] Social security department data matrix , where n2 is the sample number and d2 is the feature number;

[0066] The original data matrix is dimensionally reduced through an encoder to generate a feature vector of the original data matrix of each department, where:

[0067] ,

[0068] In the formula, Z is the feature vector of the department; X is the original data matrix; W e is the weight matrix of the encoder; b e is the bias vector; and RELU is the activation function; The symbol represents a mapping function from the original data matrix X to the virtual feature vector Z, and clearly reflects the process of converting the original data through the feature editor;

[0069] In a specific implementation, the dimension of the feature vector after dimension reduction ;

[0070] In a specific implementation, the hash value of the original data matrix is recorded in the blockchain, and the original data matrix is deleted immediately after the feature vector is generated. Different departments can only obtain a specific dimension of Z, for example, the civil affairs department cannot see the “logistics time limit”. While reducing the data dimension, the statistical rules required by the business are accurately retained, the individual sensitive information is completely stripped, and the goal of government affair collaboration that the data is available but invisible, that is, only the statistical features required for cross-department collaboration are retained, and other sensitive information is invisible, is achieved.

[0071] In a specific implementation, the original data matrix also identifies key features through a double-layer attention mechanism, including:

[0072] The local attention weight is calculated through the following formula:

[0073] ​ ,

[0074] In the formula, For the original data matrix, the first... One feature; x j For the original data matrix, the first... j One characteristic, d The total number of features in the original data matrix; Represents the original data matrix. The local attention weights of each feature, for example, in civil affairs data, the weight of the building damage level is 0.92, while the weight of the administrative division code is 0.05; W a This is the attention weight matrix;

[0075] The global attention weights are calculated using the following formula:

[0076] ,

[0077] In the formula, Represents the original data matrix Global attention weights for each feature; Represents the original data matrix. Features In one specific implementation, the cosine similarity to the target decision label Y represents historical decision results or reference standards related to business objectives, used to measure the relevance of current features to business needs. Y, as a historical decision label, can be decision results formed in similar past scenarios, such as material dispatch plans or rescue priorities. Its role is to provide a business-oriented reference for judging feature importance, ensuring that the selected features are closely related to actual government task objectives. This represents the maximum cosine similarity between all features of the original data matrix and the target decision label;

[0078] In this embodiment, the local attention weight represents the importance of the feature dimension; the global attention weight represents the correlation between the feature and the target label.

[0079] The overall weight is calculated using local attention weights and total attention weights;

[0080] In one specific implementation, the overall weight is calculated using the following formula:

[0081] ,

[0082] In the formula, Represents the original data matrix. The combined weight of each feature.

[0083] In one specific implementation, features with a comprehensive weight greater than or equal to a preset comprehensive weight threshold, such as 0.3, are identified as key features of the original data matrix. At the same time, other features, such as direct identifiers including ID card numbers and phone numbers, are automatically filtered to achieve the purpose of de-identification.

[0084] Calculate the KL divergence value to determine the degree of privacy protection of the feature vector;

[0085] In one specific implementation, the KL divergence value is calculated using the following formula:

[0086] ,

[0087] ,

[0088] In the formula, Represents the coding distribution Compared with the reference distribution KL divergence function between them; μ It is an eigenvector Z The mean vector; Σ is the eigenvector. Z The covariance matrix; is the preset KL divergence threshold; k is the dimension of the original data matrix; Represents the coding distribution; Indicates the reference distribution; Represents a normal distribution;

[0089] An n-dimensional feature vector matrix is ​​obtained by concatenating the feature vectors of each department. This feature vector matrix is ​​then combined with a rank-defining matrix of rank n-1 to obtain a joint feature matrix, where:

[0090] ,

[0091] In the formula, F Represents a joint feature matrix of multiple departments. These represent virtual feature vectors for n government departments. This indicates a row-by-row concatenation operation, which is a joint matrix formed by horizontally concatenating the feature matrices of different departments. M Let be a rank-deficient matrix, where ;

[0092] In one specific implementation, when At that time, the solution space of the system of equations has a dimension of 1, requiring at least n linearly independent equations to determine a unique solution, while the attacker can only obtain... The equation is mathematically proven to be unsolvable in reverse, meaning that the processed eigenvector matrix cannot be used to derive the original data matrix.

[0093] In a specific implementation, the joint feature matrix also performs reversibility verification, which is verified by the following formula:

[0094] ,

[0095] In the formula, represents the condition number of the matrix, and the larger the value, the more ill-conditioned the matrix, and the more unstable the inverse solution; represents the Jacobian matrix of the joint feature matrix; represents the largest singular value of the Jacobian matrix, which measures the amplification ability of the matrix to input changes in the most sensitive direction, and the larger the value, the more a small perturbation in the input will lead to a significant change in the output. represents the smallest singular value of the Jacobian matrix, which measures the amplification ability of the matrix to input changes in the least sensitive direction, and the smaller the value, the weaker the impact of input changes on the output;

[0096] When is greater than a preset value, it indicates that the joint feature matrix reversibility verification result meets the standard.

[0097] In a specific implementation, before dimensionality reduction by the encoder, the original data matrix also adjusts the dimension according to the feedback of the information loss and the privacy protection degree change, which is calculated by the following formula:

[0098] ,

[0099] In the formula, is the adjusted dimension of the original data matrix; is the dimension of the original data matrix before adjustment; represents the information loss change; is the privacy protection degree change; represents the learning rate.

[0100] In a specific implementation, the initial dimension k(0) of the original data matrix is min(d1, d2) / 2, and the upper limit k max is min(d1, d2)×0.6.

[0101] In a specific implementation, the system is built-in with a government feature dictionary to ensure that the features after dimensionality reduction are interpretable.

[0102] By constructing a loss function for balancing the privacy protection degree and information loss of the original data, the feature vector and joint feature matrix that minimize the loss function value are determined,

[0103] In a specific implementation, the loss function is calculated by the following formula:

[0104] ,

[0105] wherein, L is a loss function; Y is a target decision label; W d is a decoder weight matrix corresponding to the joint feature matrix; represents information loss, i.e. the error between the decoded data matrix and the output target; λ is a privacy protection intensity coefficient; γ is a feature sparsification coefficient; represents an encoder weight matrix W e L1 norm of the L1 norm is calculated in a manner of sum of absolute values of all elements in the matrix, which functions to realize feature sparsification, and through sparsity constraint on the weight matrix, prompts the model to preferentially retain features more critical to the task, reduces redundant information, thereby improving the explainability and computational efficiency of the features while ensuring the performance of the model.

[0106] In a specific implementation, the parameters including the encoder weight matrix, the decoder weight matrix and the bias vector are also adjusted by the gradient descent method, and the adjustment is performed through the following formula:

[0107] ;

[0108] wherein, is the adjusted parameter; is the parameter before adjustment; is a learning rate; is an effect gain, in a specific implementation, the effect gain refers to the improvement amplitude of the task performance such as the material scheduling accuracy of the model under the current parameter adjustment, and is used to measure the improvement degree of the parameter update on the task effect. The task performance such as the prediction error and the decision accuracy before and after the parameter update can be compared to determine the effect gain. When <5%, it means that the current parameter adjustment has limited improvement on the task effect, at this time, the gain coefficient ξ is triggered to participate in the parameter update, so as to further optimize the model parameters; represents the gradient of the loss function; represents the gain coefficient;

[0109] The joint feature matrix is sent to multiple departments participating in government affair collaboration, so that the multiple departments make collaborative decisions according to the joint feature matrix.

[0110] In a specific implementation, the collaborative decision determines the best decision according to the joint decision probability distribution of the joint feature matrix, and the calculation is performed through the following formula:

[0111] ;

[0112] In the formula, S is a probability distribution after the joint feature matrix is mapped by the decision weight matrix; W s is a decision weight matrix, and the dimension is m is the number of decision categories, and k is the dimension of the joint feature matrix; in a specific implementation, the softmax function can convert the result after the joint feature matrix F is mapped by the weight matrix into a probability distribution. Specifically, the softmax function performs exponentialization processing on the input value and normalization, so that each element in the output decision vector S is a probability value in the interval [0, 1], and the sum of all elements is 1. This characteristic enables S to intuitively reflect the possibility of different decision schemes, and the optimal decision scheme such as the best material scheduling scheme can be determined by selecting the element with the maximum probability, that is, S*=arg max (S), which meets the needs of decision explicitness and executability in the government affairs scene;

[0113] The decision with the maximum probability is obtained through the probability distribution as the best decision.

[0114] In a specific implementation, the best decision also comprehensively scores the decision effect in combination with the degree of privacy protection, and is calculated through the following formula:

[0115] ,

[0116] In the formula, represents the decision effect score, represents the prediction accuracy; is the time consumption of a traditional decision; is the time consumption of the best decision; in a specific implementation, the time consumption, as a part of the effect score, is closely related to the target decision. For example, the government affairs scene such as disaster emergency and material scheduling has extremely high requirements for timeliness, and flood rescue needs to meet the demand of golden 72-hour response. Traditional manual coordination or complex calculation methods are difficult to meet the requirements due to the long time consumption. One of the core goals of the embodiment is to improve the timeliness of emergency response, and the calculation complexity of cross-department cooperation is reduced through the federated computing mechanism to realize rapid response. Therefore, the time consumption, as a key indicator for measuring the efficiency of the scheme, is included in the comprehensive verification system to comprehensively evaluate the technical effect, not only to ensure the prediction accuracy and privacy security, but also to verify the advantage of the time efficiency, so as to ensure the feasibility of the scheme in the actual government affairs scene; is the KL divergence value;

[0117] When the decision effect score is greater than a preset score threshold, it indicates that the effect of the best decision is better than that of the traditional decision.

[0118] Embodiment two:

[0119] Figure 2 is a structural schematic diagram of an electronic device in an embodiment of the present application, likeFigure 2 As shown in the figure, the apparatus includes a processor 201, a memory 202, a bus 203, and a computer program stored in the memory 202 and executable on the processor 201, the processor 201 includes one or more processing cores, the memory 202 is connected to the processor 201 through the bus 203, the memory 202 is used for storing program instructions, and the processor executes the computer program to implement the steps in the above method embodiments of the embodiment one of the application.

[0120] Further, as an executable solution, the electronic device can be a computer unit, which can be a desktop computer, a notebook, a palm computer, a cloud server, and the like. The computer unit can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above-mentioned constituent structure of the computer unit is only an example of the computer unit, and does not constitute a limitation on the computer unit, and can include more or fewer components than the above, or combine certain components, or different components. For example, the computer unit can also include an input / output device, a network access device, a bus, and the like, and the embodiments of the application do not limit this.

[0121] Further, as an executable solution, the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, and the like. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor and the like, and the processor is the control center of the computer unit, and connects all parts of the computer unit through various interfaces and lines.

[0122] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the computer unit by running or executing the computer program and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; the data storage area can store data created according to the use of the mobile phone and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0123] Embodiment three:

[0124] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the method in the above-mentioned embodiments.

[0125] The modules / units integrated by the computer unit, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes of the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can realize the steps of each method embodiment when being executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM) and software distribution medium, etc. It should be noted that the computer readable medium can include or exclude contents according to the requirements of legislation and patent practice in the jurisdiction.

[0126] Embodiment four:

[0127] The application further provides a computer program product, which includes a computer program, and the computer program is executed by a processor to realize the steps of the multi-department government affair cooperation method as described above.

[0128] While this application has been particularly shown and described with references to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application as defined by the appended claims.

Claims

1. A multi-department government affair collaboration method, characterized in that, The method comprises the following steps: obtaining an original data matrix input by a plurality of departments participating in government collaboration; dimension reduction of the original data matrix by an encoder to generate a feature vector of the original data matrix of each department, wherein: ， where Z is the department-specific eigenvector; X is the original data matrix; W e is the encoder weight matrix; b e is the bias vector; RELU is the activation function; represents the mapping function from the original data matrix X to the virtual eigenvector Z. calculating the KL divergence value of the feature vector to determine the privacy protection degree of the feature vector; the KL divergence value is calculated by the following formula: ， ， wherein, represents a KL divergence function between the encoding distribution and the reference distribution ; μ is a mean vector of the feature vector Z ;∑ is a covariance matrix of the feature vector Z ; is a preset KL divergence threshold; k is a dimension of the original data matrix; represents an encoding distribution; represents a reference distribution; represents a normal distribution; splicing the feature vectors of each department to obtain an n-dimensional feature vector matrix, combining the feature vector matrix with a rank-deficient matrix with a preset rank of n-1 to obtain a joint feature matrix, wherein: ， wherein, F denotes a multi-department joint feature matrix, denotes a virtual feature vector of the n government departments, respectively, denotes a row splicing operation, i.e., a joint matrix formed by horizontally splicing the feature matrices of different departments; M is a rank-deficient matrix, wherein ​ determining the feature vector and the joint feature matrix that minimize the loss function value by constructing a loss function for balancing the privacy protection degree and information loss of the original data, the loss function being calculated by the following formula: ， In the formula, L is a loss function; Y is a target decision label; W d is a decoder weight matrix corresponding to the joint feature matrix; represents information loss; λ is a privacy protection intensity coefficient; γ is a feature sparsification coefficient; represents an encoder weight matrix W e L1 norm of sending the joint feature matrix to the plurality of departments participating in government collaboration for collaborative decision-making by the plurality of departments based on the joint feature matrix. 2.The multi-department government affair collaboration method of claim 1, wherein, The original data matrix also identifies key features through a double-layer attention mechanism, comprising: calculating the local attention weight by the following formula: ， In the formula, For the original data matrix, the first... One feature; x j For the original data matrix, the first... j One characteristic, d The total number of features in the original data matrix; Represents the original data matrix. Local attention weights for each feature; W a This is the attention weight matrix; calculating the global attention weight by the following formula: ， In the formula, Represents the original data matrix. Global attention weights for each feature; Represents the original data matrix. Features Cosine similarity to the target decision label Y; This represents the maximum cosine similarity between all features of the original data matrix and the target decision label; calculating the comprehensive weight by the local attention weight and the overall attention weight, identifying features with a comprehensive weight greater than or equal to a preset comprehensive weight threshold as key features of the original data matrix, the comprehensive weight being calculated by the following formula: ， wherein represents the original data matrix of the first characteristics. 3.The multi-department government affair collaboration method of claim 1, wherein, Before dimension reduction by the encoder, the original data matrix also adjusts the dimension according to the feedback of the information loss and the privacy protection degree variation, which is calculated by the following formula: ， In the formula, is the dimension of the adjusted original data matrix; is the dimension of the original data matrix before adjustment; represents the change in information loss; is the change in the degree of privacy protection; represents the learning rate. 4.The multi-department government affair collaboration method of claim 1, wherein, The joint feature matrix also performs reversibility verification, which is verified by the following formula: ， wherein denotes the matrix condition number; denotes the Jacobian matrix of the joint feature matrix; denotes the largest singular value of the Jacobian matrix; denotes the smallest singular value of the Jacobian matrix; When If the value is greater than the preset value, it indicates that the joint feature matrix reversibility verification result meets the standard.

5. The multi-department government affair collaboration method of claim 1, wherein, The parameters including the encoder weight matrix, the decoder weight matrix and the bias vector are also adjusted by the gradient descent method, which is adjusted by the following formula: ； wherein is the adjusted parameter; is the unadjusted parameter; is the learning rate; is the effect gain; denotes the gradient of the loss function; denotes the gain coefficient.

6. The multi-department government affair collaboration method of claim 1, wherein, The collaborative decision-making determines the best decision according to the joint decision probability distribution of the joint feature matrix, which is calculated by the following formula: ； In the formula, S is a probability distribution after the joint feature matrix is mapped by the decision weight matrix; W s is a decision weight matrix; The decision with the maximum probability obtained by the probability distribution is taken as the best decision.

7. The multi-department government affair collaboration method of claim 6, wherein, The best decision also comprehensively scores the decision effect in combination with the privacy protection degree, which is calculated by the following formula: ， wherein, denotes the decision effect score, denotes the prediction accuracy; is the time consumption of the traditional decision; is the time consumption of the optimal decision; is the KL divergence value; When the decision effect score is greater than a preset score threshold, it indicates that the best decision effect is better than the traditional decision.

8. An electronic device, comprising: The memory stores at least one program, and the processor executes the at least one program to implement the multi-department government collaboration method of any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores at least one program, and the processor executes the at least one program to implement the multi-department government collaboration method of any one of claims 1 to 7.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the multi-department government collaboration method of any one of claims 1 to 7.

Images