Feature mapping table data sharing privacy protection method and system using distribution constraint, equipment and medium
By constructing perturbation samples and Gaussian noise generators, and combining feature encoder-decoder adversarial training with Wasserstein distance constraints, the data availability problem of generative adversarial networks when tabular data is insufficient is solved, achieving privacy protection and data consistency.
Patent Information
- Application Number
- CN202511036921.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2025-11-07
AI Technical Summary
Existing data sharing privacy protection methods based on generative adversarial networks suffer from low data availability because the distribution of the generated shareable data differs significantly from the original data distribution when the original tabular data is insufficient.
By acquiring raw privacy data to construct perturbation samples, using a feature encoder to extract perturbation feature vectors, reconstructing the original data, and optimizing the feature encoder and decoder through adversarial training, a shareable feature vector is generated by combining Gaussian distributed random sampling noise, the Wasserstein distance is calculated to constrain the data distribution consistency, and the shareable data is output.
Effectively retain key and sensitive information, improve the usability of generated data, ensure that the statistical characteristics of shared data are consistent with those of the original data, and avoid data bias caused by privacy protection.
Smart Images
Figure CN120910906A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data privacy security, and particularly relates to a feature mapping table data sharing privacy protection method, system, device and medium using distribution constraint. BACKGROUND
[0002] The data sharing privacy protection method achieves privacy protection by generating shareable data to replace original data. The generated shareable data has no explicit corresponding relationship with the original data, but can show similar results to the original data in downstream tasks, thereby protecting the privacy of the original data from being leaked while reducing the loss of data use value. The data sharing privacy protection method has been widely applied to information collection, data release, privacy calculation and feature mining tasks. The data sharing privacy protection technology based on a generative adversarial network is the current mainstream research direction, which can be divided into two methods: original data-based shareable data generation and data enhancement-based shareable data generation.
[0003] The original data-based shareable data generation method directly uses original data to train a generative adversarial network, so that the generator can learn the latent distribution of the data, and the quality of the generated data can be improved by optimizing the network structure or loss function. However, this method usually relies on large-scale original data. In the case of insufficient data, the discriminator is prone to overfitting, which leads to a decline in model training effect and reduces the usability of the shareable data. The data enhancement-based shareable data generation method first enhances the original data, and then uses the enhanced samples to train a generative adversarial network. After training, the data generated by the generator is mapped back to the original data space before enhancement. In the scene of table data processing, a table data enhancement method based on a mask autoencoder is often used. This method generates perturbation samples by randomly masking part of the features, and uses an encoder to extract feature enhancement representation, and then reconstructs the original data through a feature decoder. However, the use of mean square error loss function for data reconstruction has weak ability to capture data distribution information, and using the feature representation of enhanced data for generative adversarial network training will reduce the usability of the generated data.
[0004] In summary, the existing generative adversarial network-based data sharing privacy protection method usually relies on a large amount of original data. When dealing with a small amount of table data, the method using data enhancement uses mean square error loss to constrain data reconstruction, which easily leads to the distribution of the generated shareable data deviating from the distribution of the original data, affecting the usability of the shareable data. SUMMARY
[0005] In view of the above existing problems, the present application is proposed. Therefore, the present application provides a feature mapping table data sharing privacy protection method using distribution constraint to solve the problem that the distribution of the shareable data generated by the existing method is greatly different from the distribution of the original data when the original table data is insufficient, resulting in low data usability.
[0006] To solve the above technical problems, the present application provides the following technical solutions:
[0007] In a first aspect, the present application provides a feature mapping table data sharing privacy protection method using distribution constraints, comprising: obtaining original privacy data, constructing a perturbed sample according to the original privacy data, and extracting a perturbed feature vector using a feature encoder;
[0008] Based on the perturbed feature vector, the original privacy data is reconstructed by a feature decoder, and the feature encoder and decoder are optimized by adversarial training;
[0009] A generator is constructed by randomly sampling noise based on a Gaussian distribution to obtain a shareable feature vector;
[0010] Based on the optimized feature encoder and decoder, the shareable feature vector is mapped to the original data space, the Wasserstein distance between the original data and the reconstructed data sensitive attribute distribution is calculated, and if the distance is less than a threshold value, the shareable data is output.
[0011] As a preferred scheme of the feature mapping table data sharing privacy protection method using distribution constraints, wherein: constructing a perturbed sample according to the original privacy data and extracting a perturbed feature vector using a feature encoder comprises:
[0012] The original privacy data is perturbed by a random mask mechanism to obtain a perturbed sample;
[0013] The feature of the perturbed sample is extracted using a feature encoder to obtain a perturbed feature vector.
[0014] As a preferred scheme of the feature mapping table data sharing privacy protection method using distribution constraints, wherein: reconstructing the original privacy data by a feature decoder and optimizing the feature encoder and decoder by adversarial training comprises:
[0015] The perturbed feature vector is reconstructed into privacy data by a feature decoder, the reconstruction error of the feature decoder is calculated as a data reconstruction loss function, and the feature encoder and decoder are optimized by a stochastic gradient descent algorithm;
[0016] A data discriminator is constructed to discriminate between the original data and the decoder generated data, and to perform adversarial training with the feature encoder and decoder, and the feature encoder and decoder are iteratively optimized by a stochastic gradient descent algorithm until a preset convergence condition is met to obtain the optimized feature encoder and decoder.
[0017] As a preferred scheme of the feature mapping table data sharing privacy protection method using distribution constraints, wherein: constructing a generator by randomly sampling noise based on a Gaussian distribution to obtain a shareable feature vector comprises:
[0018] The noise is randomly sampled from a Gaussian distribution, the generator obtains the feature representation, and a shareable feature vector is obtained;
[0019] The feature discriminator is constructed, the shareable feature vector and the perturbed feature vector are discriminated, and the generator is trained in an adversarial manner, the feature discriminator and the generator are iteratively optimized by a stochastic gradient descent algorithm until a preset stopping condition is reached, and the optimized generator is output;
[0020] The beneficial effect of the preferred embodiment is that the feature vector is generated by random sampling of Gaussian noise, which can be mapped one-to-one from the original data and avoid direct leakage.
[0021] As a preferred embodiment of the feature mapping table data sharing privacy protection method using distribution constraint, the output shareable data comprises:
[0022] The noise is randomly sampled from a Gaussian distribution, the generator obtains the feature representation, and a shareable feature vector is obtained;
[0023] The Wasserstein distance of the sensitive attribute distribution in the original privacy data and the reconstructed privacy data is calculated, and if the distance is less than a set threshold, the reconstructed privacy data is output as shareable data;
[0024] The beneficial effect of the preferred embodiment is that the sensitive attribute distribution difference is quantified by the Wasserstein distance, ensuring that the shared data is consistent with the statistical characteristics of the original data, and avoiding data deviation caused by privacy protection.
[0025] As a preferred embodiment of the feature mapping table data sharing privacy protection method using distribution constraint, the data discriminator is constructed, comprising:
[0026] The data discriminator C is constructed d , the Wasserstein distance of the privacy data distribution reconstructed by the feature decoder D f and the original data distribution is calculated as the loss function of the data discriminator is expressed as:
[0027]
[0028] The loss function of the feature decoder is expressed as:
[0029]
[0030] wherein, is sampled from the data output by the feature decoder, denotes sampling from the original data, denotes sampling from the data distribution and the data distribution uniformly sampling on the straight line between the sampling points, λ denotes a hyperparameter adjusting the degree of gradient constraint.
[0031] As a preferred scheme of the feature mapping table data sharing privacy protection method using distribution constraint provided by the application, wherein: the Wasserstein distance between the original privacy data sensitive attribute distribution P and the reconstructed privacy data sensitive attribute distribution Q is calculated, including:
[0032] For continuous attributes, the calculation formula is:
[0033]
[0034] Where D[P,Q] represents the Wasserstein distance between probability distributions P and Q, p j represents the probability of the jth interval value appearing in the probability distribution P, q j represents the probability of the jth interval value appearing in the probability distribution Q, and m represents the number of interval divisions.
[0035] For discrete attributes, the calculation formula is:
[0036]
[0037] Where D[P,Q] represents the Wasserstein distance between probability distributions P and Q, p i represents the probability of the ith discrete value appearing in the probability distribution P, q i represents the probability of the ith discrete value appearing in the probability distribution Q, and m represents the total number of discrete values in the distribution.
[0038] In a second aspect, the application provides a feature mapping table data sharing privacy protection system using distribution constraint, comprising:
[0039] An extraction module is configured to obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder.
[0040] An optimization module is configured to reconstruct the original privacy data based on the perturbation feature vector through a feature decoder, and optimize the feature encoder and decoder through adversarial training.
[0041] An acquisition module is configured to randomly sample noise through a Gaussian distribution, and construct a generator to obtain a shareable feature vector.
[0042] An output module is configured to map the shareable feature vector to the original data space based on the optimized feature encoder and decoder, calculate the Wasserstein distance between the original data and the sensitive attribute distribution of the reconstructed data, and output the shareable data when the Wasserstein distance is less than a threshold.
[0043] In a third aspect, the present application provides a computer device, comprising:
[0044] a memory and a processor;
[0045] The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, so as to implement the steps of the feature mapping table data sharing privacy protection method with distribution constraint.
[0046] In a fourth aspect, the present application provides a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions are executed by a processor to implement the steps of the feature mapping table data sharing privacy protection method with distribution constraint.
[0047] Compared with the prior art, the present application has the following beneficial effects: the present application calculates the Wasserstein distance between the original privacy data and the sensitive attribute distribution of the reconstructed privacy data, so that the key sensitive information can be effectively preserved and controlled, and the usability of the generated data is improved; the data is enhanced by generating a perturbation sample, and a data discriminator is constructed, and the feature encoder and decoder are trained in an adversarial manner, so as to constrain the distribution of the reconstructed privacy data to be consistent with that of the original privacy data. BRIEF DESCRIPTION OF DRAWINGS
[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.
[0049] Figure 1 The whole flowchart of the feature mapping table data sharing privacy protection method with distribution constraint according to an embodiment of the present application is shown in the figure.
[0050] Figure 2 The principle diagram of the feature mapping table data sharing privacy protection method with distribution constraint according to an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION
[0051] In order to make the above objectives, characteristics and advantages of the present application more apparent, more comprehensible, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.
[0052] Embodiment 1
[0053] With reference to Figures 1-2 For an embodiment of the present application, a feature mapping table data sharing privacy protection method using distribution constraint is provided, comprising:
[0054] S101, obtaining original privacy data, constructing a perturbation sample according to the original privacy data, and extracting a perturbation feature vector using a feature encoder;
[0055] S102, based on the perturbation feature vector, reconstructing the original privacy data through a feature decoder, and optimizing the feature encoder and decoder through adversarial training;
[0056] S103, constructing a generator to obtain a shareable feature vector by randomly sampling noise in a Gaussian distribution;
[0057] S104, based on the optimized feature encoder and decoder, mapping the shareable feature vector to the original data space, calculating the Wasserstein distance of the sensitive attribute distribution of the original data and the reconstructed data, and outputting the shareable data if the distance is less than a threshold.
[0058] In a preferred embodiment, constructing a perturbation sample according to the original privacy data and extracting a perturbation feature vector using a feature encoder comprises:
[0059] The original privacy data is perturbed by a random mask mechanism to obtain a perturbation sample;
[0060] The feature encoder is used to extract the features of the perturbation sample to obtain a perturbation feature vector.
[0061] Specifically, the original privacy data x is perturbed by a random mask m to obtain a perturbation sample The perturbation process is:
[0062]
[0063] Wherein, represents the data sampled from the empirical distribution of each feature in the original data.
[0064] It should be noted that the feature encoder maps the perturbed sample to a low-dimensional feature space by adaptive network structure and privacy constraint training, extracts the perturbed sample feature, and obtains a perturbed feature vector, wherein the adaptive network structure can adopt a fully connected neural network, a convolutional neural network or a graph neural network.
[0065] In an alternative embodiment, the feature encoder can be an adaptive network structure of a fully connected neural network, and for structured table data, the low-dimensional feature vector is output through the input layer, a plurality of hidden layers (ReLU / Sigmoid activation function), and the perturbed feature vector is extracted, wherein the hidden layer can add Dropout or BatchNorm to improve the generalization ability.
[0066] In another alternative embodiment, the feature encoder can be an adaptive network structure of a convolutional neural network, and for image or gridded data, local features are extracted through a convolutional layer, dimensionality reduction is performed through a pooling layer, and a feature vector is output through a fully connected layer to extract a perturbed feature vector.
[0067] It should be noted that by constructing a perturbed sample, the risk of exposing sensitive information of the original data can be directly reduced, the feature encoder extracts a low-dimensional perturbed feature vector, retains the useful pattern of the data while reducing redundant information, and provides an efficient representation for subsequent processing.
[0068] In a preferred embodiment, the original privacy data is reconstructed by the feature decoder, and the feature encoder and the feature decoder are optimized by adversarial training, including:
[0069] The perturbed feature vector is reconstructed into privacy data by the feature decoder, the reconstruction error of the feature decoder is calculated as a data reconstruction loss function, and the feature encoder and the feature decoder are optimized by a stochastic gradient descent algorithm;
[0070] A data discriminator is constructed to distinguish between the original data and the decoder generated data, and to perform adversarial training with the feature encoder and the decoder, and the feature encoder and the decoder are iteratively optimized by a stochastic gradient descent algorithm until a preset convergence condition is met, to obtain an optimized feature encoder and decoder.
[0071] In a preferred embodiment, the data discriminator is constructed, including:
[0072] The data discriminator C is constructed d The Wasserstein distance between the privacy data distribution reconstructed by the feature decoder D f and the original data distribution is calculated as the loss function of the data discriminator is represented as:
[0073]
[0074] Meanwhile, a loss function of the feature decoder is calculated is represented as:
[0075]
[0076] wherein, represents sampling from the data output by the feature decoder, represents sampling from the original data, represents sampling from the data distribution and the data distribution are uniformly sampled on the straight line between the sampling points, λ represents a hyperparameter for adjusting the degree of gradient constraint, and λ = 10 is usually taken; the data discriminator and the feature decoder are iteratively updated using a stochastic gradient descent algorithm until a preset convergence condition is met, that is, the loss function and the variation amplitudes in the last 5 iterations are all lower than a preset convergence threshold 10 -3 or the maximum training round number 500 is reached, and the optimized feature encoder and decoder are obtained.
[0077] In an optional embodiment, the data reconstruction loss function selects a mean square error (MSE), which minimizes the difference between the decoder reconstructed data and the original data, is sensitive to outliers, is suitable for Gaussian distribution data, and is simple to implement, stable in gradient, and the specific formula of the loss function is as follows:
[0078]
[0079] wherein, n represents the number of samples, X i the i-th sample of the original private data, represents the i-th sample of the decoder reconstructed data.
[0080] In another optional embodiment, the data reconstruction loss function selects a cross-entropy loss, which measures the difference between the true distribution and the model prediction distribution, directly optimizes the probability distribution, ensures that the class proportion of the generated data is consistent with the original data, is suitable for the case where the original private data contains classification features, the last layer of the decoder uses a Sigmoid (binary classification) or Softmax (multi-classification) activation function, the generated data retains the statistical distribution of the classification features, and the specific formula of the loss function is as follows:
[0081]
[0082] It should be noted that the original data is reconstructed by the decoder, which can ensure that the feature vector retains key information and maintains data usability, and the adversarial training minimizes the reconstruction error while protecting privacy, balancing privacy and data usability.
[0083] In a preferred embodiment, the generator obtains the sharable feature vector by constructing a Gaussian distribution random sampling noise, comprising:
[0084] The generator obtains the sharable feature vector by constructing a Gaussian distribution random sampling noise;
[0085] The feature discriminator is constructed to discriminate the sharable feature vector and the perturbed feature vector, and is trained against the generator, and the feature discriminator and the generator are iteratively optimized by a stochastic gradient descent algorithm until a preset stopping condition is reached, and the optimized generator is output.
[0086] In the present embodiment, the feature discriminator C f is trained against the generator G, and the loss function l G of the generator is calculated, and the calculation formula is:
[0087]
[0088] The loss function l f is calculated according to the distribution difference between the perturbed feature vector and the sharable feature vector generated by the random sampling noise, and is expressed as:
[0089]
[0090] wherein, represents the feature representation from the generator construction, is the feature representation extracted by the encoder, is sampled from the data distribution and the data distribution P G The straight line between the sampling points is uniformly sampled, μ represents a hyperparameter for adjusting the gradient constraint degree, and usually μ = 10, the feature discriminator and the generator are iteratively optimized by a stochastic gradient descent algorithm until a preset stopping condition is reached, i.e. the loss function l G and l f The change amplitude in the last 5 iterations is less than a preset convergence threshold 10 -3 or reaches a maximum training round 500 rounds, and the optimized generator is obtained.
[0091] It should be noted that the feature vector is generated by random sampling of Gaussian noise, which can completely break away from one-to-one mapping of original data, avoid direct leakage, and generate diversified features to meet the sharing demand while ensuring the authenticity of generated data through distribution constraint.
[0092] In a preferred embodiment, the output sharable data comprises:
[0093] Randomly sample noise from a Gaussian distribution, input the optimized generator to obtain a shareable feature vector, and use the optimized feature decoder to reconstruct the shareable feature vector into privacy data;
[0094] The Wasserstein distance of the sensitive attribute distribution of the original privacy data and the reconstructed privacy data is calculated, and if the distance is less than a set threshold, the reconstructed privacy data is output as shareable data.
[0095] In a preferred embodiment, the Wasserstein distance of the sensitive attribute distribution P of the original privacy data and the sensitive attribute distribution Q of the reconstructed privacy data includes:
[0096] For continuous attributes, the calculation formula is:
[0097]
[0098] where D[P, Q] represents the Wasserstein distance between probability distributions P and Q, p j represents the probability of the jth interval value appearing in the probability distribution P, q j represents the probability of the jth interval value appearing in the probability distribution Q, and m represents the number of intervals of the distribution.
[0099] For discrete attributes, the calculation formula is:
[0100]
[0101] where D[P, Q] represents the Wasserstein distance between probability distributions P and Q, p i represents the probability of the ith discrete value appearing in the probability distribution P, q i represents the probability of the ith discrete value appearing in the probability distribution Q, and m represents the total number of discrete values in the distribution.
[0102] If the distance is less than a set threshold of 0.01, the reconstructed privacy data is output as shareable data.
[0103] It should be noted that the present application calculates the Wasserstein distance of the sensitive attribute distribution of the original privacy data and the reconstructed privacy data, so that the key sensitive information is effectively retained and controlled, and the usability of the generated data is improved; by generating perturbation samples for data enhancement, a data discriminator is constructed, and the feature encoder and decoder are trained in an adversarial manner to constrain the distribution of the reconstructed privacy data to be consistent with the original privacy data.
[0104] The above is a schematic scheme of the feature mapping table data sharing privacy protection method using distribution constraints of the present embodiment. It should be noted that the technical scheme of the feature mapping table data sharing privacy protection system using distribution constraints is the same as the technical scheme of the feature mapping table data sharing privacy protection method using distribution constraints described above. The technical details of the feature mapping table data sharing privacy protection system using distribution constraints in the present embodiment are not described in detail, and can be referred to the description of the technical scheme of the feature mapping table data sharing privacy protection method using distribution constraints described above.
[0105] Embodiment 2
[0106] The present embodiment provides a feature mapping table data sharing privacy protection system using distribution constraints, comprising:
[0107] The extraction module is configured to obtain original privacy data, construct a perturbed sample according to the original privacy data, and extract a perturbed feature vector using a feature encoder.
[0108] The optimization module is configured to reconstruct the original privacy data based on the perturbed feature vector through a feature decoder, and optimize the feature encoder and decoder through adversarial training.
[0109] The acquisition module is configured to obtain a shareable feature vector by constructing a generator through random sampling of noise in a Gaussian distribution.
[0110] The output module is configured to map the shareable feature vector to the original data space based on the optimized feature encoder and decoder, calculate the Wasserstein distance of the sensitive attribute distribution of the original data and the reconstructed data, and output the shareable data if the distance is less than a threshold.
[0111] The present embodiment also provides a computer device suitable for the case of feature mapping table data sharing privacy protection using distribution constraints, comprising:
[0112] The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement the feature mapping table data sharing privacy protection method using distribution constraints as proposed in the above embodiment.
[0113] The present embodiment also provides a storage medium having a computer program stored thereon, which is executed by a processor to implement the feature mapping table data sharing privacy protection method using distribution constraints as proposed in the above embodiment.
[0114] The storage medium proposed in the present embodiment and the feature mapping table data sharing privacy protection method using distribution constraints proposed in the above embodiment belong to the same inventive concept. The technical details not described in detail in the present embodiment can be referred to the above embodiment, and the present embodiment has the same beneficial effects as the above embodiment.
[0115] Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be realized by software and necessary general hardware, and of course can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a FLASH, a hard disk, or an optical disc, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of various embodiments of the present application.
[0116] Embodiment 3
[0117] Referring to Table 1, a feature mapping table data sharing privacy protection method using distribution constraints is provided for an embodiment of the present application. In order to verify its beneficial effects, economic benefit calculation and simulation experiments are carried out for scientific demonstration.
[0118] The experimental data selected two commonly used handwritten digit recognition data sets Mnist12 and Mnist28 to verify the availability of the shared data generated by the method in the downstream prediction task, and the experimental data is shown in Table 1.
[0119] Table 1: Details of experimental data sets
[0120] Dataset Training set / test set B M Task type Mnist12 60k / 10k 144 1 Multi-classification Mnist28 60k / 10k 784 1 Multi-classification
[0121] Wherein, B represents the number of attributes with 2-class discrete characteristics in the data set; M represents the number of attributes with multi-class discrete characteristics in the data set, and the last column is the type of machine learning task applicable to the data set. 5% and 10% of the original training set are used as the training set to simulate different data volume missing scenarios.
[0122] In the process of the experiment, the number of encoder-decoder hidden layers is set to 2, the number of generator hidden layers is set to 3, the number of discriminator hidden layers is set to 3, and the feature representation dimension is set to 32.
[0123] The accuracy (Accuracy) is used to evaluate the availability of the shared data in the experiment, and the accuracy calculation method is represented as:
[0124]
[0125] Wherein, TP is the number of positive class samples predicted as positive class, FN is the number of positive class samples predicted as negative class, FP is the number of negative class samples predicted as positive class, and TN is the number of negative class samples predicted as negative class.
[0126] The experiment is performed on a computer and a server, the specific configuration of the computer is: Inter i9-9900, CPU 3.10GHz, memory 32G, and the operating system is windows 10, 64-bit; the specific configuration of the server is: NVIDIARTX 600024G, and the operating system is Linux Ubuntu 64-bit.
[0127] According to the step method of the present application, the test results are obtained: the experiment uses the distributed constraint feature mapping table data sharing privacy protection method to evaluate the usability of Mnist12 corpus and Mnist28 corpus, the prediction accuracy of the present application is 58% and 63.1% under the 5% and 10% training data missing scenarios on Mnist12 corpus, and the prediction accuracy is 64.8% and 68.5% under the 5% and 10% training data missing scenarios on Mnist28 corpus, and the present application has good effect under different data missing scenarios.
[0128] It should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.
Claims
1. A privacy protection method for data sharing using a distributed constraint feature mapping table, characterized by, The method comprises the following steps: obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; based on the perturbation feature vector, reconstruct the original privacy data through a feature decoder, and optimize the feature encoder and decoder through adversarial training; obtain a shareable feature vector by constructing a generator through random sampling of noise from a Gaussian distribution; based on the optimized feature encoder and decoder, map the shareable feature vector to the original data space, calculate the Wasserstein distance between the sensitive attribute distribution of the original data and the reconstructed data, and output the shareable data if the distance is less than a threshold. 2.The privacy protection method for feature mapping table data sharing with distribution constraint according to claim 1, wherein, The method comprises the following steps: obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; 3.The privacy protection method for feature mapping table data sharing with distribution constraint according to claim 1, wherein, The method comprises the following steps: reconstruct the original privacy data through a feature decoder, and optimize the feature encoder and decoder through adversarial training; obtain a shareable feature vector by constructing a generator through random sampling of noise from a Gaussian distribution; 4.The privacy protection method for feature mapping table data sharing with distribution constraint according to claim 3, wherein, based on the optimized feature encoder and decoder, map the shareable feature vector to the original data space, calculate the Wasserstein distance between the sensitive attribute distribution of the original data and the reconstructed data, and output the shareable data if the distance is less than a threshold. The method comprises the following steps: obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; 5.The privacy protection method for feature mapping table data sharing with distribution constraint according to claim 4, wherein, obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; The method comprises the following steps: reconstruct the original privacy data through a feature decoder, and optimize the feature encoder and decoder through adversarial training; 6.The privacy protection method for feature mapping table data sharing with distribution constraint according to claim 3, wherein, obtain a shareable feature vector by constructing a generator through random sampling of noise from a Gaussian distribution; constructing a data discriminator C d , computing a feature decoder D f reconstructing the privacy data distribution and the original data distribution as a loss function of the data discriminator is represented as: Loss function of the feature decoder is represented as: wherein, denotes sampling from the data output by the feature decoder, denotes sampling from the original data, denotes sampling from the data distribution and the data distribution uniformly sample points on the straight line between the sampling points, λ denotes a hyperparameter that adjusts the degree of gradient constraint. 7.The privacy protection method for feature mapping table data sharing with distribution constraint according to claim 5, wherein, based on the optimized feature encoder and decoder, map the shareable feature vector to the original data space, calculate the Wasserstein distance between the sensitive attribute distribution of the original data and the reconstructed data, and output the shareable data if the distance is less than a threshold. The method comprises the following steps: where D[P, Q] denotes the Wasserstein distance between probability distributions P and Q, p j denotes the probability of the jth interval taking a value in the probability distribution P, q j denotes the probability of the jth interval taking a value in the probability distribution Q, and m denotes the number of intervals of the distribution partition; obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; where D[P, Q] denotes the Wasserstein distance between probability distributions P and Q, p i denotes the probability of the ith discrete value occurring in probability distribution P, q i denotes the probability of the ith discrete value occurring in probability distribution Q, and m denotes the total number of discrete values in the distribution.
8. A privacy protection system for data sharing using distribution-constrained feature mapping table, applying the privacy protection method for data sharing using distribution-constrained feature mapping table according to any one of claims 1 to 7, characterized in that, obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; The method comprises the following steps: reconstruct the original privacy data through a feature decoder, and optimize the feature encoder and decoder through adversarial training; obtain a shareable feature vector by constructing a generator through random sampling of noise from a Gaussian distribution; based on the optimized feature encoder and decoder, map the shareable feature vector to the original data space, calculate the Wasserstein distance between the sensitive attribute distribution of the original data and the reconstructed data, and output the shareable data if the distance is less than a threshold. The method comprises the following steps: obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; obtain original privacy data, construct a perturbation sample according to the original privacy data, and extract a perturbation feature vector using a feature encoder; The method comprises the following steps: reconstruct the original privacy data through a feature decoder, and optimize the feature encoder and decoder through adversarial training; obtain a shareable feature vector by constructing a generator through random sampling of noise from a Gaussian distribution; based on the optimized feature encoder and decoder, map the shareable feature vector to the original data space, calculate the Wasserstein distance between the sensitive attribute distribution of the original data and the reconstructed data, and output the shareable data if the distance is less than a threshold. An output module is configured to map the shareable feature vector to the original data space based on the optimized feature codec, calculate a Wasserstein distance between the original data and the sensitive attribute distribution of the reconstructed data, and output the shareable data if the Wasserstein distance is less than a threshold.
9. A computer device, comprising: Comprising: a memory and a processor; the memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, and the computer executable instructions, when executed by the processor, implement the steps of the feature mapping table data sharing privacy protection method with distribution constraints according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer executable instructions, when executed by the processor, implement the steps of the feature mapping table data sharing privacy protection method with distribution constraints according to any one of claims 1 to 7.