A trusted table data generation method and related equipment for protecting multi-party differential privacy

Abstract

This invention discloses a method and related equipment for generating trusted tabular data that protects the differentiated privacy of multiple parties. The method includes: receiving local statistical information from local tabular data to obtain global statistical information; obtaining the aggregation weights of each client based on the global statistical information; distributing the current global model parameters to each client to obtain updated local model parameters; constructing a target global model based on the aggregation weights and updated local model parameters; generating initial privacy data through the target global model, and filtering the initial privacy data based on preset negation constraints to obtain trusted privacy data. This invention can generate tabular data while protecting the privacy of multiple parties and improving the trustworthiness of the tabular data, and can be widely applied in the fields of federated learning and data generation technology.

Description

Technical Field

[0001] This invention relates to the fields of federated learning and data generation technology, and in particular to a trusted tabular data generation method and related equipment that protects the privacy of multiple parties with differentiating characteristics. Background Technology

[0002] With the increasing demand for high-quality data, federated learning has expanded from a simple model training paradigm to the field of data generation. It aims to address the problems of data scarcity and data silos by generating data consistent with the real distribution through multi-party collaboration. Early research primarily focused on adapting generative adversarial networks (GANs) or variational autoencoders (VAEs) to federated frameworks. One approach utilizes AT-GANs to address class imbalance in structured tabular data. However, this AT-GAN approach focuses primarily on statistical distribution fitting, neglecting the strict constraints between tabular data features, potentially leading to generated minority class samples that violate domain common sense. Additionally, there is the VAE-BGM synthetic data sharing strategy, which generates data locally and shares it among nodes, effectively improving the model adaptability of low-resource nodes. However, this method of directly sharing synthetic data still carries the risk of leaking the original data when differential privacy protection is weak.

[0003] To further enhance privacy and stability, federated GANs are currently used to address the issue of incomplete data distribution for single clients. However, in practical applications, model training may fail to converge, resulting in a lack of diversity in the generated data. Existing techniques also include introducing Wasserstein distance and K-Lipschitz constraints into FedDPGAN to stabilize training, combined with differential privacy mechanisms to prevent gradient leakage. However, the random noise introduced by differential privacy can easily disrupt cross-feature constraints in tabular data. Furthermore, while the proposed FedDPGAN mitigates the impact of non-independent and identically distributed data through novel loss functions and outlier removal strategies, it focuses on unsupervised cross-modal retrieval tasks and is not specifically optimized for cross-feature constraints in high-dimensional tabular data.

[0004] Furthermore, in distributed scenarios, the data held by the participants often exhibits non-independent and identically distributed characteristics. Non-independent and identically distributed data often has multifaceted impacts on federated learning. This type of data not only directly or indirectly affects model convergence and training efficiency but also reduces the communication efficiency among the participants. Summary of the Invention

[0005] In view of this, the main objective of the embodiments of the present invention is to provide a reliable table data generation method and related equipment that protects the differentiated privacy of multiple parties, in order to solve at least one of the problems of the prior art. The present invention can generate table data while protecting the privacy data of multiple parties and improve the reliability of table data.

[0006] To achieve the above objectives, one aspect of the present invention provides a method for generating trusted tabular data that protects the differentiated privacy of multiple parties, the method comprising: Receive local statistics from local table data to obtain global statistics; Based on the global statistics, obtain the aggregate weight of each client; The current global model parameters are distributed to each client to obtain the updated local model parameters. Based on the aggregated weights and the updated local model parameters, a target global model is constructed. Initial privacy data is generated through the target global model, and then filtered based on a preset negative constraint to obtain trusted privacy data.

[0007] In some embodiments, obtaining the aggregate weight of each client based on the global statistics includes the following steps: Based on the column data type of the global statistics, obtain the difference measurement value between the global statistics and the local statistics; Based on the difference measurement value, the number of clients, and the number of local data columns of the clients, a first divergence matrix is ​​constructed; Divide each element in the first scatter matrix by the sum of the elements in its column to obtain the second scatter matrix; The second divergence matrix is ​​summed by row operations to obtain a third divergence matrix; the elements of the third divergence matrix are the deviation metric scores of the client. Obtain the initial global data distribution ratio, set an exponential scaling factor on the initial global data distribution ratio, and obtain the target global data distribution ratio; The deviation metric score is normalized, and the complement of the normalized deviation metric score is obtained. The complement is combined with the target global data distribution ratio to obtain the comprehensive deviation score. The overall deviation score is passed to the softmax function to obtain the aggregate weight.

[0008] In some embodiments, obtaining the difference measurement value between the global statistics and the local statistics based on the column data type of the global statistics includes the following steps: When the column data type is discrete, the difference between the global statistics and the local statistics is measured by JS divergence to obtain the JS divergence distance value; When the column data type is continuous, the Wasserstein distance value is obtained by measuring the difference between the global statistics and the local statistics using the Wasserstein distance metric. The difference measurement includes the JS divergence distance value and the Wasserstein distance value.

[0009] In some embodiments, the step of sending the current global model parameters to each client and obtaining the updated local model parameters includes the following steps: The local table data is standardized using the client to obtain a real sample; The current global model parameters are obtained through the client; the current global model parameters include the current generator parameters and the current discriminator parameters. Random noise is sampled by the client and input into the generator to generate pseudo samples; The client performs proportional interpolation between the real sample and the pseudo sample to obtain the interpolated sample. Obtain the gradient norm of the interpolated sample and construct a gradient penalty term whose gradient norm is close to 1; Based on the real samples, the pseudo samples, and the gradient penalty term, a discriminator loss function is constructed, and the current discriminator parameters are updated to obtain the updated discriminator parameters. The generator gradient is obtained through the client, and the generator gradient is clipped and noise is introduced to obtain a noisy gradient. The current generator parameters are updated based on the noise gradient to obtain the updated generator parameters; The updated discriminator parameters and the updated generator parameters are uploaded to the server through the client, and the step of obtaining the current global model parameters through the client is returned until a preset training round is reached or the target global model converges, and the updated local model parameters are obtained; the updated local model parameters include the updated discriminator parameters and the updated generator parameters.

[0010] In some embodiments, constructing the target global model based on the aggregated weights and the updated local model parameters includes the following steps: Based on the aggregate weights, the updated local model parameters are weighted and summed to obtain the initial global model; The initial global model is subjected to a noise-adding operation to obtain the target global model.

[0011] In some embodiments, generating initial privacy data through the target global model and filtering the initial privacy data based on a preset negative constraint to obtain trusted privacy data includes the following steps: The preset negation constraint is parsed to obtain multiple sub-constraints; the sub-constraints are contained within the preset negation constraint; Iterate through the initial privacy data and select any two pieces of the initial privacy data as two pieces of data to be inspected. When two pieces of data to be inspected satisfy any of the sub-constraints, the two pieces of initial privacy data corresponding to the two pieces of data to be inspected are removed to obtain trusted privacy data.

[0012] To achieve the above objectives, another aspect of the present invention provides a trusted table data generation apparatus for protecting the differential privacy of multiple parties, the apparatus comprising: The information acquisition module is used to receive local statistical information from local table data and obtain global statistical information; The weight acquisition module is used to acquire the aggregate weight of each client based on the global statistical information; The parameter acquisition module is used to send the current global model parameters to each of the clients and obtain the updated local model parameters. The model building module is used to build a target global model based on the aggregated weights and the updated local model parameters; The data generation module is used to generate initial privacy data through the target global model, and filter the initial privacy data based on a preset negative constraint to obtain trusted privacy data.

[0013] To achieve the above objectives, another aspect of the present invention provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described above.

[0014] To achieve the above objectives, another aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0015] To achieve the above objectives, another aspect of the present invention provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions to cause the computer device to perform the aforementioned method.

[0016] The embodiments of the present invention include at least the following beneficial effects: The present invention provides a method and related equipment for generating trusted tabular data that protects the differentiated privacy of multiple parties. This scheme receives local statistical information from local tabular data and integrates it to obtain global statistical information, laying a reliable data foundation for subsequent analysis; based on the global statistical information, it obtains the aggregate weights of each client, realizing a quantifiable evaluation of contribution; it distributes the current global model parameters to each client and obtains the updated local model parameters, realizing distributed collaborative training, while better adapting to the data characteristics of different clients; based on the aggregate weights and the updated local model parameters, it effectively integrates the training results of each client to form a more comprehensive target global model; it generates initial privacy data through the target global model, and filters the initial privacy data based on a fast credibility enhancement mechanism of preset negation constraints, which can quickly discover and remove data that violates the constraints, thereby improving the credibility of the synthesized data and generating trusted privacy data while protecting the privacy data of multiple parties. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a trusted table data generation method for protecting the differential privacy of multiple parties, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the trusted table data generation framework (FedPriGen) for protecting the differential privacy of multiple parties provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Privacy-Fidelity balance curves under different noise levels provided in the embodiments of the present invention; Figure 4 This is a schematic diagram illustrating the differences in the correlation between data generated by different models and real data features provided in this embodiment of the invention; Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0020] It should be noted that although functional modules are divided in the system diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first / S100" and "second / S200" in the specification, claims, and the foregoing drawings may be used herein to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are used only to distinguish one concept from another. For example, first information may also be referred to as second information without departing from the scope of the embodiments of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the words "if" or "when" as used herein may be interpreted as "when," "in response to a determination," or "in the event of a determination."

[0021] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0023] Before providing a detailed description of the embodiments of the present invention, some of the nouns and terms involved in the embodiments of the present invention will be explained first. The nouns and terms involved in the embodiments of the present invention are subject to the following interpretations.

[0024] Jensen-Shannon Divergence (JS divergence) is an index that measures the similarity between two probability distributions. It is a symmetric improvement of KL divergence and has excellent properties such as nonnegativity, symmetry, and boundedness.

[0025] Wasserstein distance (WD) is a measure of the difference between two probability distributions. It originates from optimal transport theory and is commonly used in machine learning and statistics.

[0026] Existing methods for generating data in federated scenarios pay little attention to the data quality degradation caused by introduced differential privacy noise, failing to guarantee the credibility of data generated by the global model. When dealing with non-independent and distributed data, pseudo-data is generated locally on multiple clients to balance the distribution differences between clients, but the quality of the generated pseudo-data is not addressed. Furthermore, the modeling capabilities for data held locally by participating parties are insufficient, making it unable to fit complex data distributions. Moreover, existing methods do not address the credibility detection methods for generated data, nor the ease of operation of credibility detection.

[0027] In view of this, this invention provides a method and related device for generating trustworthy tabular data that protects the differential privacy of multiple parties. To mitigate the global distribution shift caused by non-independent and identically distributed data, this solution designs a global consistency calculation method and a method for setting aggregation weights. This reduces the negative impact of participants with significant deviations from the global distribution on the model, thereby statistically improving the consistency between the generated data and the real joint data distribution. Furthermore, based on the integration of a differential privacy module during multi-party collaborative training, this solution introduces a filtering mechanism based on negation constraint rules at the generation end. This mechanism can quickly identify data that violates the constraints and remove it from the generated dataset, thereby improving the overall credibility of the synthesized tabular data. Moreover, it can generate trustworthy tabular data that conforms to specific domain rules while protecting the privacy of multiple parties.

[0028] Figure 1 This is an optional flowchart of a trusted table data generation method for protecting the differentiated privacy of multiple parties provided in an embodiment of the present invention. Figure 1 The method may include, but is not limited to, steps S100 to S500: Step S100: Receive local statistical information from local table data to obtain global statistical information; Step S200: Obtain the aggregate weight of each client based on global statistics; Step S300: Distribute the current global model parameters to each client and obtain the updated local model parameters; Step S400: Construct the target global model based on the aggregated weights and the updated local model parameters; In step S500, initial privacy data is generated through the target global model, and the initial privacy data is filtered based on preset negative constraints to obtain trusted privacy data.

[0029] In step S100 of some embodiments, the local client needs to send local statistical information of the local table data to the server for subsequent calculation of the client's aggregation weight. The local statistical information includes the frequency distribution of attribute values ​​in each discrete column of data, and the Gaussian mixture distribution fitted to the continuous columns. The distribution information includes the mean of the fitted Gaussian distribution. Variance of Gaussian distribution and the weight values ​​of each Gaussian distribution After receiving local statistics from all local table data, the server can synthesize global statistics. During this process, the server does not directly access the local table data, but only accesses the statistical information, achieving "usable but invisible" data. This integrates global statistics while protecting the privacy of the original data of all participating parties, laying a reliable data foundation for subsequent analysis.

[0030] In some embodiments, after receiving global statistics, the server can obtain discrete data columns. Global statistics and continuous data columns Global statistics Then, based on the discrete data columns Global statistics and continuous data columns Global statistics Construct the first encoder for each discrete data column. and a second encoder for continuous data columns Subsequently, the server will use the first encoder. Second encoder Distributed to each client. The client then uses the encoder (first encoder) to... Second encoder The local table data is encoded, transforming it into standardized feature vectors. The client uses these encoded feature vectors as input to initialize (or train) the local model (generator and discriminator). The locally initialized model has the same network structure, and the discriminator and generator networks are predefined and data-independent.

[0031] In step S200 of some embodiments, after receiving statistical information from the local data of each client, the server calculates the aggregate weights for each client. These aggregate weights are used to globally aggregate the trained local model parameters (i.e., the updated local model parameters) in scenarios where they are not independent and identically distributed, laying the foundation for constructing the target global model. For example, the calculation process of the aggregate weights is as follows:

[0032] In some embodiments, step S200 may include, but is not limited to, steps S210 to S270: Step S210: Based on the column data type of the global statistics, obtain the difference measurement value between the global statistics and the local statistics; Step S220: Construct the first divergence matrix based on the difference measurement value, the number of clients, and the number of local data columns of the clients; Step S230: Divide each element in the first scatter matrix by the sum of the elements in its column to obtain the second scatter matrix; Step S240: Perform matrix row summation on the second divergence matrix to obtain the third divergence matrix; the elements of the third divergence matrix are the client's deviation metric scores. Step S250: Obtain the initial global data distribution ratio, set an exponential scaling factor on the initial global data distribution ratio, and obtain the target global data distribution ratio; Step S260: Normalize the deviation metric score and obtain the complement of the normalized deviation metric score. Combine the complement with the target global data distribution ratio to obtain the comprehensive deviation score. Step S270: Pass the overall deviation score to the softmax function to obtain the aggregate weight.

[0033] In step S210 of some embodiments, a specific difference measurement method is selected based on the column data type of the global statistics information, thereby obtaining the client... Local data column Local statistics With column Global statistics The specific method for measuring differences between columns needs to be selected based on the data type of the columns. Optionally, Jensen-Shannon Divergence (JSD) is used to measure differences between discrete data; Wasserstein Distance (WD) is used to measure differences between continuous columns.

[0034] In some embodiments, step S210 may include, but is not limited to, steps S211 to S212: Step S211: When the column data type is discrete, the JS divergence distance value is obtained by measuring the difference between global statistics and local statistics using JS divergence. Step S212: When the column data type is continuous, the Wasserstein distance value is obtained by measuring the difference between global statistics and local statistics using the Wasserstein distance metric. The difference measurements include the JS divergence distance and the Wasserstein distance.

[0035] In step S211 of some embodiments, when the column data type is discrete, JS divergence is used to measure the difference between discrete data. In JS divergence measurement, two probability vectors... and The JSD between is mathematically defined as ,in yes and The pointwise mean, It is the Kullback-Leibler divergence. The JS divergence distance is symmetric and bounded between 0 and 1. For each discrete column... and client Calculated column Global statistics With the client Regarding the list Local statistics The JS divergence distance between them, i.e. .

[0036] In step S212 of some embodiments, when the column data type is continuous, the Wasserstein distance is used to measure the difference between continuous columns. The Wasserstein distance defines two distributions. and The differences are: ; In the formula, Indicates Wasserstein distance; Denotes the infimum of a set; Describing a joint distribution specifies how the distributions are... The probability mass shift in the distribution middle; Represents the space of real pairs The set of probability distributions on the given surface, where the margins of the first and second factors are respectively... and ; Point Time The distance; Indicates in joint distribution Down, The probability infinitesimal element at the location.

[0037] Wasserstein distance can be interpreted as the minimum cost of transforming one distribution into another, where the cost is the number of distributions to be moved multiplied by the distance that must be moved. For each client... Each consecutive column in Difference measurement value For the client Middle data column Statistical information With column Global statistics Wasserstein distance between them.

[0038] In step S220 of some embodiments, before performing aggregate weight calculation, it is first necessary to construct a The first divergence matrix ,in It is the number of clients. This represents the number of local data columns on the client side. First scatter matrix. Each element in Representative client Local data column Local statistics With column Global statistics The difference between them. For example, the expression for the first scatter matrix is: ; In the formula, , , , These are elements in the first scatter matrix; For the client's index; This is an index for the client's local data column.

[0039] In step S230 of some embodiments, the first divergence matrix After construction is complete, the first scatter matrix is... According to each client In each column Normalization is then performed. For example, each element in the first scatter matrix is ​​divided by the sum of the elements in its corresponding column to obtain the second scatter matrix, calculated as follows: ; This step ensures that all columns have the same weight range (0 to 1, with the sum of each column element equal to 1) while maintaining the relative differences between different clients and the global column data distribution.

[0040] In step S240 of some embodiments, different columns in the second scatter matrix are aggregated by performing matrix row summation on the second scatter matrix. The divergence between the two sides is used to obtain the third divergence matrix, which is calculated using the following formula: ; For each client The obtained deviation measurement score It accounts for the differences between client-side and global distributions, but it does not yet take into account the differences between local table data on the client side.

[0041] In step S250 of some embodiments, the initial global data distribution ratio is obtained. , This would result in clients with large amounts of local table data having a higher weight. To effectively control this excessive weighting and to incentivize local clients with smaller local table data volumes but whose data distribution is similar to the global distribution, an exponential scaling factor was introduced. This allows us to obtain the target global data distribution ratio. .when This will increase the weight of clients with large amounts of data, therefore The value is usually less than 1, such as 0.7-0.9.

[0042] In step S260 of some embodiments, the deviation measurement score for all clients is first calculated. Normalization to interval, then use complement To represent distribution consistency, and to compare the distribution ratio of locally available data to global data. By combining these data, the differences in data values ​​and amounts from each client are integrated, ultimately resulting in a comprehensive deviation score. This comprehensive deviation score takes into account differences in data volume and data distribution. For example, the formula for calculating the difference between each client's data value and data volume is as follows: .

[0043] In step S270 of some embodiments, the comprehensive deviation score is... The result is passed to the softmax function, which can calculate the value for each client. Aggregate weights The calculation formula is as follows: ; When the server aggregates the global model, the client... Aggregate weights.

[0044] In steps S300 to S400 of some embodiments, the server distributes the current global model parameters to each local client. The local client uses the current global model parameters as initial parameters to train its local model (generator and discriminator). During training, to protect privacy, gradient clipping and noise addition are performed on the local model parameters. Subsequently, the updated local model parameters with noise are uploaded to the server. The server aggregates the updated local model parameters according to the calculated aggregation weights to form new current global model parameters, which are then redistributed to each client until a certain number of training rounds are reached or the global model converges, at which point training ends.

[0045] In some embodiments, step S300 may include, but is not limited to, steps S310 to S390: Step S310: Standardize the local table data using the client to obtain a real sample; Step S320: Obtain the current global model parameters through the client; the current global model parameters include the current generator parameters and the current discriminator parameters; Step S330: Random noise is sampled through the client and input into the generator to generate pseudo samples; Step S340: The client performs proportional interpolation between the real samples and the pseudo samples to obtain the interpolated samples. Step S350: Obtain the gradient norm of the interpolated sample and construct a gradient penalty term whose gradient norm is close to 1. Step S360: Based on the real samples, pseudo samples, and gradient penalty term, construct the discriminator loss function and update the current discriminator parameters to obtain the updated discriminator parameters; Step S370: Obtain the generator gradient through the client, trim the generator gradient and introduce noise to obtain the noisy gradient; Step S380: Update the current generator parameters according to the noise gradient to obtain the updated generator parameters; Step S390: Upload the updated discriminator parameters and updated generator parameters to the server through the client, and return the step of obtaining the current global model parameters through the client until the preset training rounds are reached or the target global model converges, and obtain the updated local model parameters; the updated local model parameters include the updated discriminator parameters and the updated generator parameters.

[0046] In step S310 of some embodiments, local table data is used as training samples. The client performs standardization processing on the local table data, thereby transforming the local table data into a multidimensional vector. To obtain real samples This facilitates subsequent model calculations; at the same time, the generator in the client maps noise to the representation space where the real samples are located, which helps the generator learn the potential real distribution of the real samples.

[0047] In step S320 of some embodiments, the client receives the current global model parameters sent by the server and uses these parameters as initial parameters for training the generator network and discriminator network. The received current global model parameters include the current generator parameters. and current discriminator parameters .

[0048] In step S330 of some embodiments, random noise is sampled and input into a generator to generate pseudo-samples. Exemplarily, from the distribution of random noise... In the process, a noise vector is randomly sampled. The sampled random noise vector Input generator The output is calculated by the generator network. The result is a pseudo-sample. .

[0049] In step S340 of some embodiments, linear interpolation is performed between the real samples on the local client and the generated pseudo samples to obtain interpolated samples. For example, the expression for the interpolated samples is: ; In the formula, Indicates the interpolated sample; Represents a real sample; Indicates a pseudo-sample; The interpolation coefficient is a scalar value that is uniformly randomly sampled between 0 and 1, and determines the weight of real samples and generated samples (pseudo samples) in the mixed sample. Represents each interpolated sample The values ​​are all selected independently and randomly.

[0050] In step S350 of some embodiments, the gradient of the interpolated sample is calculated, and the gradient is obtained. Norm, which forces the gradient to be norm A norm approaching 1 yields a regularized gradient penalty term, which constrains the magnitude of model parameter updates. For example, this gradient penalty term can be expressed as... .in, This represents a hyperparameter used to control the weights of the gradient penalty term; This represents the expectation of the interpolated sample; This indicates the distribution of the interpolated samples; Discriminator The gradient of the interpolated sample; express Norm.

[0051] In step S360 of some embodiments, a discriminator loss function can be constructed based on real samples, pseudo samples, and a gradient penalty term, and the current discriminator parameters are updated to obtain the updated discriminator parameters. By introducing a gradient penalty term into the discriminator loss function, the instability and mode collapse problems during model training on the local client can be alleviated, thereby stabilizing the training process and improving the diversity of generated pseudo data samples. For example, the discriminator loss function expression after adding the gradient penalty term is: ; In the formula, This represents the discriminator loss value; This represents an expectation of the actual data; Represents the distribution of the real samples; This represents the expectation for pseudo-samples; This represents the distribution of pseudo-samples.

[0052] In step S370 of some embodiments, considering the importance of user privacy, a sampling Gaussian mechanism (SGM) is used to prune the gradient learned by the generator and introduce noise to enhance privacy. For example, the generator gradient is obtained as follows: ; Next, order The generator gradient is clipped using the following formula: ; Then, noise is introduced into the cropped generator gradient, and the calculation formula is as follows: ; In the formula, , Represents the generator gradient; Indicates the size of the small batch; Represents the number of samples collected from the distribution of random noise. A random noise vector; Indicates that the generator is based on The generated first A spurious sample; Discriminator For the Output score of each pseudo-sample; This represents the generator gradient after clipping; Indicates the gradient with added noise; Indicates the noise level; This represents a standard Gaussian distribution.

[0053] In step S380 of some embodiments, the current generator parameters are updated based on the noise gradient and by minimizing the generator loss function to obtain the updated generator parameters. The calculation formula includes: ; ; In the formula, This represents the generator loss value; This represents the expectation for pseudo-samples; Represents the distribution of random noise; The generator uses noise Generated pseudo-samples; Indicates the updated generator parameters; Indicates the current generator parameters; This represents the generator's learning rate. By minimizing the generator's loss function and updating the generator parameters, the generator can be prompted to produce pseudo-samples that more closely resemble the real sample data.

[0054] In step S390 of some embodiments, after the local client has completed training, the client will update the local model parameters. The updated discriminator and generator parameters (including the updated parameters) are uploaded to the server. The server then performs global model aggregation based on a distribution-consistent aggregation scheme, using the local model parameters uploaded by the client as the new current global model parameters. Noise is added to the aggregated global model. If the noisy global model fails to converge, and the client's local model has not yet reached the preset training epochs, the process returns to obtaining the current global model parameters from the client. This continues until the preset training epochs are reached or the target global model converges, at which point training ends.

[0055] In some embodiments, step S400 may include, but is not limited to, steps S410 to S420: Step S410: Based on the aggregate weights, perform a weighted summation of the updated local model parameters to obtain the initial global model; Step S420: Add noise to the initial global model to obtain the target global model.

[0056] In step S410 of some embodiments, after the local client has completed training, the server updates the local model parameters uploaded by the client. The global model is aggregated using the client's aggregated weights; that is, the updated local model parameters are weighted and summed to obtain the initial global model. For example, the expression for the initial global model is: ; In the formula, This represents the initial global model.

[0057] In step S420 of some embodiments, the target global model is obtained by adding noise to the initial global model. For example, the expression of the target global model is: ; In the formula, Represents the global model of the target; This represents the noise vector added to the target global model; Represents the identity matrix.

[0058] In step S500 of some embodiments, a target global model is used. Generate initial privacy data The generated initial privacy data is filtered using a fine-grained checking algorithm based on negation constraints (as shown in Table 1) to finally obtain trustworthy privacy data. And thus obtain A set of data, i.e., a reliable dataset. .

[0059] A denial constraint (DC) describes the state relationships and schema design between attributes. Denial constraints are introduced as a generalization of many other dependencies. A DC can express the attribute relationships specified by these dependencies, as well as many other attribute relationships beyond them. For example, consider the relation instances shown in Table 2. : (1) Functional dependency (FD): It can be declared that a tuple that has the same value on attribute A must also have the same value on attribute B; (2) Unique Column Combination (UCC): It can be declared that the combined value of attribute A and attribute C is unique throughout the table (i.e., there are no two different tuples with exactly the same combination of values ​​in A and C). (3) Order dependency (OD): It can be declared that for any two tuples, the tuple with the larger value on attribute C must also have a larger value on attribute D.

[0060] Table 1 Fine-grained inspection algorithm

[0061] Table 2 Relationship Examples

[0062] Using the notation for negative constraints, all attribute relationships in Table 2 can be specified using the following expression: (1) ; (2) ; (3) .

[0063] Negation constraints have a strong dependency on the dataset, and qualified data can be filtered out based on the constraints. Assuming that the dependencies between data have been obtained in advance according to the specific task (these dependencies need to be defined in combination with the actual situation), then the generated data can be filtered according to the dependencies, thereby obtaining data that is more in line with the real distribution and more practical for the specific task.

[0064] In some embodiments, step S500 may include, but is not limited to, steps S510 to S530: Step S510: Parse the preset negation constraint to obtain multiple sub-constraints; the sub-constraints are contained within the preset negation constraint. Step S520: Traverse the initial privacy data and select any two initial privacy data as two data to be inspected; Step S530: When two pieces of data to be inspected satisfy any sub-constraint, the two initial privacy data corresponding to the two pieces of data to be inspected are removed to obtain trusted privacy data.

[0065] In step S510 of some embodiments, the predefined negation constraint is... The analysis reveals multiple sub-constraints. These sub-constraints are contained in These sub-constraints represent relationships that should not exist in tabular data.

[0066] In some embodiments, steps S520 to S530 involve filtering the generated initial privacy data pairwise, treating these two initial privacy data sets as two data sets to be inspected, and then determining if these two data sets satisfy any sub-constraint. If the result is false, it means that these two data entries do not meet expectations and need to be removed. At this point, the mask for these two data entries will be set to False. After all data filtering is complete, only the data with a mask of True is the expected data, i.e., reliable private data.

[0067] Negation constraint The logic is that if any one of the sub-constraints of the negation constraint is satisfied, the constraint rule is violated. During pairwise filtering of data, the two data pairs should not have any sub-constraint relationship. If the two data pairs do not satisfy any sub-constraint, they are considered qualified data.

[0068] like Figure 2As shown, this embodiment of the invention provides a trusted table data generation framework (FedPriGen) for protecting the differential privacy of multiple parties. Within this framework, a privacy-preserving trusted table data generation algorithm is applied (as shown in Table 3), and the specific process for generating trusted table data to protect the differential privacy of multiple parties is as follows: Table 3. Privacy-preserving trusted tabular data generation algorithms

[0069] Step 1: The local client needs to send local statistics of its local table data to the server for calculating client weights. After receiving the global statistics, the server obtains the discrete columns. Global statistics and global statistics for continuous columns Then, based on this information, encoders are constructed for the discrete data columns. and encoders for continuous data columns Subsequently, the server will send the encoder. and The data is distributed to each client. The client encodes its local table data using the encoder and initializes its local model. During this process, the server does not directly access the client's local table data, but only accesses the data's statistical information.

[0070] Step 2: After receiving local statistical information from the local tables of each client, the server calculates the weights for each client. The local client uses the current global model parameters sent by the server as initial parameters for local model training. During training, to protect privacy, gradient clipping and noise addition are performed on the local model parameters, and then the noisy parameters are uploaded to the server. The server aggregates the model parameters according to the calculated weights and then redistributes them to each client until a certain number of training epochs are reached or the global model converges, at which point training ends.

[0071] Step 3: After obtaining the global model, the privacy data generated by the global model is filtered using a negative constraint checking algorithm to remove data that does not conform to the constraint rules and retain data that does conform to the rules, thereby obtaining trustworthy privacy data.

[0072] In some embodiments, a comparative experiment was conducted between the proposed trusted table data generation framework (FedPriGen) for protecting the differential privacy of multiple parties and existing models / algorithms. The experimental setup is as follows: (1) Dataset: The FedPriGen model was tested on five publicly available datasets from the UCI Machine Learning Database: Adult, Philadelphia, Obesity Survey, Paddy, and Steel Industry. The experiment was conducted in a federated environment with three clients. To simulate non-independent and identically distributed data, a Dirichlet distribution was used to split the dataset, and the size of the local data for each client was unpredictable. The training and test sets were split in a 7:3 ratio, with strict isolation between the two sets to ensure data objectivity. Due to computational resource limitations, all data used in the experiment was randomly sampled from the original dataset, with each set containing 16K rows. Detailed descriptions of the experimental datasets are shown in Table 4.

[0073] Table 4. Dataset Introduction

[0074] (2) Comparison model and baseline setting: To systematically verify the superior performance of the FedPriGen model framework in generating privacy-preserving and trustworthy data in a distributed environment, this experiment selected four representative aggregation algorithms and one generative model for comparison. The four aggregation algorithms were used to replace the distribution-consistency-based aggregation algorithm in the FedPriGen model framework. Other environmental settings, such as differential privacy mechanisms, noise levels, training epochs, and the non-independent and identically distributed nature of local client data, remained the same as in the FedPriGen model framework. The names of the comparative experimental models are shown in Table 5.

[0075] Table 5 Experimental Methods

[0076] The FedAvg aggregation algorithm achieves a good balance between communication cost and model by periodically aggregating local updates from the client. However, it often suffers from non-independent and identically distributed (i.i.d.) data. The high communication cost of the existing FedSGD aggregation algorithm stems from uploading the updated local model parameters to the server after each update. While this alleviates data heterogeneity to some extent, the communication cost is too high. The FedProx aggregation algorithm adds a proximate term to the local objective function to limit excessive deviation from the global model, thereby improving optimization consistency across clients. The SCAFFOLD aggregation algorithm effectively reduces the "drift effect" caused by offset by maintaining and correcting control discrepancies between the client and server, exhibiting more stable and faster convergence performance under i.i.d. data conditions. The FedTabDiff model, in distributed scenarios, utilizes the generative capabilities of a probability diffusion model to generate privacy-preserving tabular data. The FedPriGen proposed in this invention is a data generation framework for federated scenarios that integrates aggregation methods based on distribution consistency and data inspection.

[0077] (3) Performance indicators and evaluation methods: A set of standard evaluation metrics are used to measure model performance, including: Fidelity, RangeCoverage, Privacy, Utility, Average Jensen-Shannon divergence (Avg-JSD), and Average Wasserstein distance (Avg-WD). These metrics represent different aspects of data generation quality, ensuring a comprehensive evaluation of model performance.

[0078] Fidelity assesses how well the generated data simulates real data, considering both column-level and row-level comparisons. For column fidelity, the similarity between corresponding columns in the generated and real datasets is evaluated. Numerical attributes are compared using the Kolmogorov-Smirnov Statistic (KSS) to fit empirical distributions, denoted as... Total Variance Distance (TVD) quantifies the difference in categorical attributes, expressed as... ,in Represented as attributes Medium category The probability. Fidelity calculated column-wise. The following formula can be used for calculation: ; The overall fidelity of the columns in the generated dataset is across all attributes. The average value.

[0079] Privacy metrics are used to quantify the degree to which generated data inhibits the identification of the original data. The Distance to Closest Records (DCR) is used to measure the privacy of the synthesized data; the DCR is defined as the distance to the nearest record (DCR) of the generated data point. To real data points The minimum distance is formally expressed as follows: ; in, This represents the distance metric, which in this paper uses Euclidean distance. The final score is the median of the DCR distances for all generated data points.

[0080] Range coverage is a crucial metric for assessing how well a generated data column replicates the categorical diversity of a real data column. When applied to categorical features, this metric initially identifies the unique class in the original column. The quantity. Then, it evaluates whether these categories are in the generated column. The metric calculates the proportion of the true category in the generated data, i.e. For numerical features, where and These represent the real column and the generated column, respectively. This metric is evaluated... The range (its maximum and minimum values) and How close are their ranges? This consistency is quantified by the following formula: ; Utility measures the degree of functional equivalence to the original data. This utility is quantified by training a machine learning model on the generated dataset and then evaluating its performance on the original dataset. Utility is defined as the performance of the generated dataset on the original dataset. The performance of the classifier trained on the dataset has the same dimensionality as the real training set, but was subsequently tested using the original dataset. An evaluation was conducted. This process assessed the effectiveness of the synthetic data in replicating the statistical properties required for accurate model training. The average accuracy of all classifiers was calculated to represent the overall usability of the synthetic data, formalized as follows: ; In the formula, Represents utility score, Indicates the first The accuracy of each classifier was evaluated. For a comprehensive assessment, this study selected... There are several classifiers, namely: Random Forest, Decision Trees, Logistic Regression, AdaBoost, and Naive Bayes.

[0081] The Average Jensen-Shannon Divergence (Avg-JSD) is an evaluation metric applied to categorical columns. First, the JSD distance between the generated data and the real data for each categorical column is calculated. Second, the resulting JSDs are averaged to obtain a simple and easy-to-understand score, simply called Avg-JSD. The closer the Avg-JSD is to 0, the more realistic the synthetic data is.

[0082] The Average Wasserstein Distance (Avg-WD) is an evaluation metric applied to continuous columns. Unlike the JSD, WD is unbounded and can vary significantly depending on the data size. To make WD scores comparable across columns, each continuous column in the real data is fitted and normalized before calculating WD, and the same normalization method is applied to the corresponding columns in the synthetic data. The WD scores of all columns are then averaged to obtain the final score, simply referred to as Avg-WD. The closer Avg-WD is to 0, the more realistic the synthetic data is.

[0083] I. Performance Testing The comparative experiments were first conducted on the Philadelphia dataset, comparing the proposed model against five other models or algorithms to test its performance in various aspects. Each experimental metric was the average of 10 runs conducted under the same model parameters and noise level. "This indicates that the closer the result is to 1, the better the effect." "" indicates that the closer the result is to 0, the better the effect. Specifically, the experimental results obtained on five different datasets are shown in Tables 6, 7, 8, 9, and 10.

[0084] Experimental results on five different datasets demonstrate that the FedPriGen model exhibits a more balanced and stable performance advantage across most datasets and evaluation metrics. On these five datasets, the FedPriGen model achieves relatively high scores on RangeCoverage and Fidelity, indicating that the data generated by the model can effectively cover the distribution domain of the real data while maintaining high statistical consistency. For example, on the Philadelphia and Adult datasets, the FedPriGen model scores 0.757 and 0.668 on Fidelity, respectively, significantly outperforming traditional aggregation methods such as FedAvg, FedSGD, and FedProx, demonstrating its stronger distribution modeling capability under complex data distributions.

[0085] Regarding privacy metrics, while the FedPriGen model did not achieve the highest privacy score on the five datasets mentioned above, its overall performance remained within a reasonable range, without sacrificing data usability for privacy improvements. In contrast, while the FedTabDiff and SCAFFOLD models achieved high privacy scores on some datasets, their performance on Fidelity, Utility, and distribution distance metrics was significantly poor, reflecting a clear imbalance between privacy protection and data quality. This phenomenon indicates that privacy-generating methods that rely solely on noise injection or adversarial mechanisms often struggle to simultaneously balance generation quality and practical usability when facing statistical heterogeneity caused by non-independent and identically distributed data in federated learning.

[0086] In terms of utility metrics, the FedPriGen model achieved the highest or second-highest performance across all five datasets. On the Philadelphia dataset, the FedPriGen model scored 0.753, an improvement of approximately 1.6% compared to the second-best method, FedProx, which scored 0.741. On the Adult dataset, the FedPriGen model scored 0.805, an improvement of approximately 11.0% compared to the second-best method, FedProx, which scored 0.725. On the Steel Industry dataset, the FedPriGen model scored 0.859, an improvement of approximately 16.7% compared to the second-best method, SCAFFOLD, which scored 0.736. On the ObesitySurvey and Paddy datasets, which have limitations related to missing values ​​and high-dimensionality, respectively, the FedPriGen model still maintained its leading position, scoring 0.547 and 0.667, respectively. On the Obesity Survey dataset, the performance improvement compared to the suboptimal FedProx method is approximately 7.0%; on the Paddy dataset, the performance improvement compared to the suboptimal FedSGD method is approximately 18.5%. This indicates that the data generated by the FedPriGen model has stronger transferability and decision support capabilities in downstream tasks, validating its effectiveness in real-world applications.

[0087] Further analysis from the perspective of distribution distance reveals that the FedPriGen model achieved the lowest or near-lowest Avg-JSD and Avg-WD values ​​on most datasets. On the Philadelphia dataset, the FedPriGen model scored 0.377 for Avg-JSD and 0.044 for Avg-WD, representing a decrease of approximately 3.9% and 13.7% compared to FedAvg. On the Adult dataset, the FedPriGen model scored 0.089 for Avg-JSD and 0.061 for Avg-WD, with the Avg-JSD score decreasing by approximately 21.2% compared to FedAvg. On the Steel Industry dataset, the FedPriGen model achieved an Avg-JSD score of 0.050, remaining at a low level. This indicates that the difference between the distribution of its generated data and the distribution of the real data is smaller than that of other comparative methods. This result not only verifies the effectiveness of the aggregation strategy based on distribution consistency in mitigating the impact of non-independent and identically distributed data, but also experimentally confirms that the posterior filtering mechanism based on negation constraints can significantly improve the credibility of synthetic data by eliminating logically inconsistent samples.

[0088] The experimental results above show that, compared with other comparative models and methods, the FedPriGen framework can generate more practical and reliable data while protecting privacy when dealing with non-independent and identically distributed data in federated learning.

[0089] Table 6 Performance comparison on the Philadelphia dataset

[0090] Table 7 Performance comparison on the Adult dataset

[0091] Table 8 Performance comparison on the Obesity Survey dataset

[0092] Table 9 Performance comparison on the Paddy dataset

[0093] Table 10 Performance comparison on the Steel Industry dataset

[0094] II. Privacy-Fidelity Curve To explore and balance the relationship between fidelity and privacy, different noise levels were set in the experiment, and the scores for Privacy and Fidelity were calculated at these noise levels. The experimental results are as follows: Figure 3 As shown. To make the curve smoother, Gaussian smoothing was applied. This experiment was conducted based on the Adult dataset. Initially, because the added noise was small, it did not significantly affect model learning. At this point, the Fidelity score increased with the Privacy score until it reached its maximum. At the maximum point, the Privacy score was 0.35 and the Fidelity score was 0.67, at which point the noise level was... Subsequently, as the noise level increased, the Privacy score also increased. However, the addition of a large amount of noise affected the model's learning ability, and the Fidelity score decreased significantly as the Privacy score increased.

[0095] III. Data Correlation Analysis Next, the privacy data generated by the FedPriGen model proposed in this invention and various comparative models are compared with real data to analyze the differences in feature dependencies between the generated privacy data and real data. Specific results are as follows: Figure 4 As shown.

[0096] Figure 4Parts (a), (c), and (e) in the figure represent the differences in the correlation between the data generated by the models FedTabDiff, FedAvg, and FedProx and the features of the real data. These three figures share a common characteristic: the differences are widely distributed and dispersed. Figure 4 The largest difference value in part (a) is 0.64, and most differences are above 0.11. Figure 4 The largest difference value in part (c) is 0.39, compared to Figure 4 Part (a), Figure 4 The differences in the distribution in part (c) have improved, with differences ranging from 0.1 to 0.39. Differences below 0.1 can be ignored. Figure 4 In part (e), the maximum difference is 0.38, a few differences are between 0.1 and 0.38, and the majority of the remaining differences are below 0.1.

[0097] Figure 4 Parts (b), (d), and (f) in the figure represent the correlation differences between the data generated by models SCAFFOLD, FedSGD, and FedPriGen and the real data features, respectively. These three figures show the same characteristic: a very small number of features show large differences, while the remaining features show small differences, exhibiting a bipolar characteristic. Figure 4 The maximum difference in part (b) is 0.91, with a few differences ranging from 0.38 to 0.57. Figure 4 The maximum difference in part (d) is 0.94, with a few differences ranging from 0.32 to 0.61. Figure 4 The maximum difference in part (f) is 0.92, with a few differences ranging from 0.31 to 0.61.

[0098] Comparative analysis revealed that the FedPriGen model retains most of the dependencies between features in the privacy-preserving data it generates. In this respect, the FedPriGen model outperforms the FedTabDiff, FedAvg, and FedProx models.

[0099] IV. Ablation Experiment Finally, ablation experiments were used to verify the effectiveness of each improvement in the model. The experimental results are based on those obtained from Adult. Table 11 shows the performance of different combinations of improvements.

[0100] Table 11 Ablation Experiment Results

[0101] Combination 1 shows no improvements, only the basic Fed+GAN model. In Combination 2, after adding an aggregation scheme based on distribution consistency, the macro-average performance of RangeCoverage, Fidelity, Utility, and Avg-JSD improved by approximately 2.68%, 2.45%, 4.88%, and 1.73%, respectively. Compared to Combination 2, after adding a fine-tuning operation, RangeCoverage, Fidelity, Utility, and Avg-JSD improved by approximately 2.05%, 6.7%, 10.12%, and 15.24%, respectively. The ablation experiments show that the addition of each component in the proposed framework effectively improves the performance of the global model. Therefore, each improvement to the FedPriGen model is meaningful.

[0102] This invention also provides a trusted table data generation apparatus for protecting the differential privacy of multiple parties, which can implement the above-mentioned trusted table data generation method for protecting the differential privacy of multiple parties. The apparatus includes: The information acquisition module is used to receive local statistical information from local table data and obtain global statistical information; The weight acquisition module is used to obtain the aggregate weight of each client based on global statistics. The parameter acquisition module is used to send the current global model parameters to each client and obtain the updated local model parameters. The model building module is used to build the target global model based on the aggregated weights and the updated local model parameters; The data generation module is used to generate initial privacy data through the target global model, and filter the initial privacy data based on preset negative constraints to obtain trusted privacy data.

[0103] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0104] This invention also provides an electronic device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including a tablet computer, an in-vehicle computer, or similar device.

[0105] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0106] refer to Figure 5 , Figure 5 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 601 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention. The memory 602 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 602 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 602 and is called and executed by the processor 601. The input / output interface 603 is used to implement information input and output; The communication interface 604 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 605 transmits information between various components of the device (e.g., processor 601, memory 602, input / output interface 603, and communication interface 604); The processor 601, memory 602, input / output interface 603, and communication interface 604 are connected to each other within the device via bus 605.

[0107] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0108] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0109] This invention also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions to cause the computer device to perform the aforementioned method.

[0110] In summary, the trusted table data generation method and related device for protecting the differentiated privacy of multiple parties according to embodiments of the present invention have the following advantages: 1. The data generated by the embodiments of the present invention is highly reliable. Without sharing the original data, it realizes a generation method that balances privacy protection and data quality, providing reliable data support for downstream tasks.

[0111] 2. The aggregation method considering distribution consistency proposed in this embodiment of the invention. This method first designs a global consistency calculation method, and then designs a method for setting aggregation weights. This reduces the impact of client data with large deviations from the overall distribution on the global model, thereby statistically improving the consistency between the generated data and the actual joint data distribution, enhancing the global generation model's ability to characterize the overall data distribution, and alleviating the global model bias problem caused by non-independent and identically distributed data.

[0112] 3. This invention provides a scalable rule-driven trusted data generation mechanism. Within the proposed FedPriGen framework, a rapid trustworthiness enhancement mechanism based on negative constraints is designed. This mechanism can quickly identify data that violates constraints and remove it from the generated dataset, thereby improving the overall trustworthiness of the synthesized tabular data. Furthermore, this framework can flexibly adapt to constraint rules in different domains, exhibiting good versatility and scalability.

[0113] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0114] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0115] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0116] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0117] Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions, and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents. The preferred embodiments of the present invention have been specifically described above, but the invention is not limited to the described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the invention, and all such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for generating trusted tabular data privacy preserving multi-party differential privacy, characterized in that, Includes the following steps: Receive local statistics from local table data to obtain global statistics; Based on the global statistics, obtain the aggregate weight of each client; The current global model parameters are distributed to each client to obtain the updated local model parameters. Based on the aggregated weights and the updated local model parameters, a target global model is constructed. Initial privacy data is generated through the target global model, and then filtered based on a preset negative constraint to obtain trusted privacy data.

2. The method of claim 1, wherein, The step of obtaining the aggregate weight of each client based on the global statistics includes the following steps: Based on the column data type of the global statistics, obtain the difference measurement value between the global statistics and the local statistics; Based on the difference measurement value, the number of clients, and the number of local data columns of the clients, a first divergence matrix is ​​constructed; Divide each element in the first scatter matrix by the sum of the elements in its column to obtain the second scatter matrix; The second divergence matrix is ​​summed by row operations to obtain a third divergence matrix; the elements of the third divergence matrix are the deviation metric scores of the client. Obtain the initial global data distribution ratio, set an exponential scaling factor on the initial global data distribution ratio, and obtain the target global data distribution ratio; The deviation metric score is normalized, and the complement of the normalized deviation metric score is obtained. The complement is combined with the target global data distribution ratio to obtain the comprehensive deviation score. The overall deviation score is passed to the softmax function to obtain the aggregate weight.

3. The method of claim 2, wherein, The step of obtaining the difference measurement value between the global statistics and the local statistics based on the column data type of the global statistics includes the following steps: When the column data type is discrete, the difference between the global statistics and the local statistics is measured by JS divergence to obtain the JS divergence distance value; When the column data type is continuous, the Wasserstein distance value is obtained by measuring the difference between the global statistics and the local statistics using the Wasserstein distance metric. The difference measurement includes the JS divergence distance value and the Wasserstein distance value.

4. The method of claim 1, wherein, The step of distributing the current global model parameters to each client and obtaining the updated local model parameters includes the following steps: The local table data is standardized using the client to obtain a real sample; The current global model parameters are obtained through the client; the current global model parameters include the current generator parameters and the current discriminator parameters. Random noise is sampled by the client and input into the generator to generate pseudo samples; The client performs proportional interpolation between the real sample and the pseudo sample to obtain the interpolated sample. Obtain the gradient norm of the interpolated sample and construct a gradient penalty term whose gradient norm is close to 1; Based on the real samples, the pseudo samples, and the gradient penalty term, a discriminator loss function is constructed, and the current discriminator parameters are updated to obtain the updated discriminator parameters. The generator gradient is obtained through the client, and the generator gradient is clipped and noise is introduced to obtain a noisy gradient. The current generator parameters are updated based on the noise gradient to obtain the updated generator parameters; The updated discriminator parameters and the updated generator parameters are uploaded to the server through the client, and the step of obtaining the current global model parameters through the client is returned until a preset training round is reached or the target global model converges, and the updated local model parameters are obtained; the updated local model parameters include the updated discriminator parameters and the updated generator parameters.

5. The method of claim 1, wherein, The step of constructing the target global model based on the aggregated weights and the updated local model parameters includes the following steps: Based on the aggregate weights, the updated local model parameters are weighted and summed to obtain the initial global model; The initial global model is subjected to a noise-adding operation to obtain the target global model.

6. The method of claim 1, wherein, The process of generating initial privacy data through the target global model and filtering the initial privacy data based on preset negation constraints to obtain trusted privacy data includes the following steps: The preset negation constraint is parsed to obtain multiple sub-constraints; the sub-constraints are contained within the preset negation constraint; Iterate through the initial privacy data and select any two pieces of the initial privacy data as two pieces of data to be inspected. When two pieces of data to be inspected satisfy any of the sub-constraints, the two pieces of initial privacy data corresponding to the two pieces of data to be inspected are removed to obtain trusted privacy data.

7. A trusted table data generation apparatus for protecting multi-party differential privacy, characterized by, include: The information acquisition module is used to receive local statistical information from local table data and obtain global statistical information; The weight acquisition module is used to acquire the aggregate weight of each client based on the global statistical information; The parameter acquisition module is used to send the current global model parameters to each of the clients and obtain the updated local model parameters. The model building module is used to build a target global model based on the aggregated weights and the updated local model parameters; The data generation module is used to generate initial privacy data through the target global model, and filter the initial privacy data based on a preset negative constraint to obtain trusted privacy data.

8. An electronic device, comprising: Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images