A data privacy protection method

Abstract

The application provides a data privacy protection method, comprising: obtaining an original data set, extracting a high-sensitivity field and calculating the correlation between the high-sensitivity field and other fields to determine a high-correlation field combination; performing intelligent segmentation and dispersed storage on the high-correlation field combination, and constructing a distributed hash table index; realizing encrypted aggregation and asynchronous update of global model parameters through a federal learning framework; constructing an adversarial sample for the high-correlation field combination, training a privacy protection defense model; performing differential privacy disturbance processing on the high-sensitivity field; calculating a correlation risk index of the high-correlation field combination in real time, and generating a correlation risk analysis report; dynamically adjusting a data segmentation strategy according to the correlation risk analysis report, and optimizing local model training sample selection; if the correlation risk index exceeds a preset safety threshold, triggering adaptive update of a differential privacy de-correlation processing flow.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to data security management for federated learning. BACKGROUND

[0002] Problem background:

[0003] In business scenarios, there are complex association characteristics between high-sensitivity fields such as ID numbers, mobile phone numbers, and other fields. By calculating the Pearson correlation coefficient matrix, it can be determined whether the correlation coefficients between these fields exceed the preset threshold, thereby identifying high-association field combinations. In order to solve the privacy leakage risk brought by high-association field combinations, the Apriori algorithm based on association rule mining needs to be used to intelligently segment these fields according to the minimum support and minimum confidence requirements. The segmented fields need to be stored in different data nodes and build a distributed hash table index to form a distributed data storage structure oriented to privacy protection. At the same time, in order to protect privacy during model training, a federated adversarial learning network needs to be constructed to realize the encrypted aggregation and asynchronous update of global model parameters through the federated learning framework. In the privacy protection of high-association field combinations, adversarial samples oriented to Member inference attacks need to be constructed, and a privacy protection defense model is trained by introducing a gradient penalty term, so that it can identify and resist privacy inference attacks based on data association. In order to further reduce the association between fields, the maximum mean difference method can be used to transform the original data features, and the Laplace mechanism can be used to generate differential privacy random noise to perturb the fields with high sensitivity. By establishing an association risk dynamic evaluation system, the association risk index of high-association field combinations in the data node is calculated in real time based on the information entropy principle, and a data association risk analysis report can be generated to quantify the privacy leakage risk level. According to the association risk analysis report, the data segmentation granularity and dimension can be dynamically adjusted, the field distribution balance index can be recalculated, and the local model training sample selection strategy can be optimized through the transfer learning method. If the association risk index exceeds the preset safety threshold, the adaptive update of the differential privacy de-association processing flow is triggered, and the optimal noise parameter combination is dynamically searched through the immune optimization algorithm to balance the data privacy protection strength and the data availability loss. Finally, through the cooperative interaction of the federated adversarial learning network and the differential privacy de-association processing flow, the risk response capability of the privacy protection defense model is continuously strengthened, the data segmentation and storage mechanism is continuously optimized, and the leakage risk of sensitive private information is minimized, ensuring the data analysis value while realizing data security sharing. SUMMARY

[0004] The present application provides data security management for federated learning, mainly including:

[0005] The original data set is obtained, high-sensitive fields are extracted, and correlations between the high-sensitive fields and other fields are calculated to determine a high-correlation field combination; the high-correlation field combination is intelligently segmented and stored in a decentralized manner, and a distributed hash table index is constructed; encrypted aggregation and asynchronous updating of global model parameters are realized through a federated learning framework; a privacy protection defense model is trained by constructing adversarial samples for the high-correlation field combination; differential privacy perturbation processing is performed on the high-sensitive fields; a correlation risk index of the high-correlation field combination is calculated in real time, and a correlation risk analysis report is generated; the data segmentation strategy is dynamically adjusted according to the correlation risk analysis report, and the selection of local model training samples is optimized; if the correlation risk index exceeds a preset safety threshold, adaptive updating of a differential privacy de-correlation processing procedure is triggered; through the collaborative interaction of a federated adversarial learning network and the differential privacy de-correlation processing procedure, the data segmentation and storage mechanism is continuously optimized. Further, the original data set is obtained and the high-correlation field combination is determined, including: identity card numbers and mobile phone numbers are extracted from the original data set as high-sensitive fields; a Pearson correlation coefficient matrix between the high-sensitive fields and other fields is calculated; and a high-correlation field combination with a correlation coefficient greater than a preset threshold is determined according to the Pearson correlation coefficient matrix. Further, the high-correlation field combination is intelligently segmented and stored in a decentralized manner, including: the Apriori algorithm is used to analyze the high-correlation field combination; the high-correlation field combination is intelligently segmented according to minimum support and minimum confidence requirements; the segmented data is stored in different data nodes; and a distributed hash table index is constructed based on the decentralized stored data. Further, encrypted aggregation and asynchronous updating of global model parameters are realized through a federated learning framework, including: a CNN convolutional neural network model parameter is initialized locally; encrypted aggregation of the CNN convolutional neural network model parameter is realized through a FederatedAveraging framework; and the CNN convolutional neural network model parameter is asynchronously updated. Further, adversarial samples are constructed for the high-correlation field combination and a privacy protection defense model is trained, including: adversarial samples for Member reasoning attacks are constructed based on the high-correlation field combination; a gradient penalty term is introduced during training; and a privacy protection defense model is trained through the adversarial samples and the gradient penalty term. Further, differential privacy perturbation processing is performed on the high-sensitive fields, including: a maximum mean difference method is used to perform feature transformation on the original data; differential privacy random noise is generated based on a Laplace mechanism; and the high-sensitive fields are perturbed using the differential privacy random noise.Further, the real-time calculation of the correlation risk index of the high correlation field combination and the generation of the correlation risk analysis report include: calculating the correlation risk index of the high correlation field combination based on the information entropy principle; generating a data correlation risk analysis report according to the correlation risk index; and taking the data correlation risk analysis report as the basis for subsequent dynamic adjustment. Further, the dynamic adjustment of the data segmentation strategy and the optimization of the local model training sample selection according to the correlation risk analysis report include: adjusting the data segmentation granularity and the segmentation dimension according to the correlation risk analysis report; recalculating the field distribution balance index; and optimizing the local model training sample selection strategy through the transfer learning method.

[0006] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0007] The application discloses a privacy protection data processing method based on federated learning. The method first extracts high-sensitive fields from an original data set and analyzes the correlation of the high-sensitive fields with other fields to determine high-correlation field combinations. Then, the field combinations are intelligently segmented and stored in a distributed manner to construct a distributed hash table index. Through a federated learning framework, encrypted aggregation and asynchronous updating of model parameters are realized, and a privacy protection defense model is constructed by training adversarial samples. The high-sensitive fields are subjected to differential privacy perturbation processing, and a correlation risk index is calculated in real time to generate an analysis report. Based on the report, the data segmentation strategy is dynamically adjusted and the training sample selection is optimized. When the risk exceeds a threshold, adaptive updating of differential privacy de-correlation processing is triggered. Through the cooperation of federated adversarial learning and differential privacy, the application continuously optimizes the data segmentation and storage mechanism, effectively protects the privacy of sensitive data, and ensures data availability. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 A flowchart of the data security management for federated learning of the application.

[0009] Figure 2 A schematic diagram of the data security management for federated learning of the application.

[0010] Figure 3 Another schematic diagram of the data security management for federated learning of the application.

[0011] Figure 4 Another schematic diagram of the data security management for federated learning of the application.

[0012] Figure 5 Another schematic diagram of the data security management for federated learning of the application.

[0013] Figure 6 Another schematic diagram of the data security management for federated learning of the application.

[0014] Figure 7 Yet another schematic diagram for data security management for federated learning of the present application.

[0015] Figure 8 Yet another schematic diagram for data security management for federated learning of the present application.

[0016] Figure 9 Yet another schematic diagram for data security management for federated learning of the present application.

[0017] Figure 10 Yet another schematic diagram for data security management for federated learning of the present application.

[0018] Figure 11 Yet another schematic diagram for data security management for federated learning of the present application.

[0019] Figure 12 Yet another schematic diagram for data security management for federated learning of the present application.

[0020] Figure 13 Yet another schematic diagram for data security management for federated learning of the present application. DETAILED DESCRIPTION

[0021] The technical solutions of the present application will be described clearly and completely below in conjunction with the embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0022] As Figures 1-13 , the data security management for federated learning of the present embodiment can specifically include:

[0023] Step S101, obtaining an original data set, extracting a high-sensitivity field and calculating the correlation between it and other fields to determine a high-correlation field combination; performing intelligent segmentation and dispersed storage on the high-correlation field combination, and constructing a distributed hash table index.

[0024] When obtaining the original dataset, first, the data source and structure need to be clarified. Assume that the dataset comes from a customer information table of a financial company, containing fields such as ID number, mobile number, name, address, and consumption record. Among them, ID number and mobile number belong to high-sensitive fields and need to be extracted first. After extracting the high-sensitive fields, the Pearson correlation coefficient matrix between them and other fields can be calculated to quantify their linear relationship. For example, the correlation coefficient between ID number and address field may be high, because the same ID number usually corresponds to the same address. If the preset correlation coefficient threshold is 0.7, then the field combination with a correlation coefficient greater than 0.7 will be determined as a high correlation field combination, such as ID number and address, mobile number and consumption record, etc. For intelligent segmentation of high correlation field combinations, the Apriori algorithm can be used. This algorithm finds frequent item sets and generates association rules by setting minimum support and minimum confidence. For example, taking ID number and address as a frequent item set, set the minimum support to 0.5 and the minimum confidence to 0.8, then only when the frequency of the simultaneous occurrence of these two fields exceeds 50% and the association strength exceeds 80%, will they be retained. Subsequently, according to these rules, the high correlation field combination is segmented into multiple subsets and stored in different data nodes. For example, ID number and address are stored in one node, and mobile number and consumption record are stored in another node. Building a distributed hash table index is to improve data query efficiency. By hashing the segmented field combination, a unique hash value is generated, and these hash values are mapped to the corresponding data nodes. For example, the combination of ID number and address is hashed to get a hash value, and then the hash value is associated with the data node that stores this combination. In this way, when querying, only the corresponding node needs to be quickly located according to the hash value, without traversing the entire dataset, thus improving the query efficiency and reducing the risk of data leakage.

[0025] Step S102, the encryption aggregation and asynchronous update of global model parameters are realized through the federated learning framework; the adversarial samples are constructed for the high correlation field combination, and the privacy protection defense model is trained.

[0026] To achieve encrypted aggregation and asynchronous updating of global model parameters through the federal learning framework, it is first necessary to understand the basic principles of federal learning. Federal learning is a distributed machine learning method in which participants train models locally rather than centralizing data on a central server. To ensure data privacy, the parameters of the global model are aggregated after local updates through encryption techniques. For example, assume there are three participants A, B, and C, each of whom trains a CNN model locally to generate model parameters W_A, W_B, and W_C. These parameters are uploaded to the central server through encryption techniques such as homomorphic encryption or secure multi-party computation, and the server performs a weighted average of the parameters without decryption to obtain the global model parameters W_global. This encrypted aggregation method ensures the privacy of participants' data while enabling collaborative training of the model. Asynchronous updating allows participants to upload model parameters without complete synchronization, improving system flexibility and efficiency. For example, participant A uploads parameters at time t1, participant B uploads parameters at time t2, and participant C uploads parameters at time t3. The server dynamically adjusts the global model based on the latest uploaded parameters without waiting for all participants to complete training. This asynchronous mechanism is particularly suitable for scenarios where participants have different computing capabilities or network conditions. To construct adversarial samples for high-correlation field combinations and train privacy protection defense models, it is first necessary to identify high-correlation fields in the data. For example, in a medical data set, the age, gender, and disease type of patients may have high correlation, and attackers may infer the identity of specific patients through these fields. To defend against such inference attacks, adversarial samples can be generated to disrupt the model's output. For example, assume a model predicts disease type based on age and gender. Adversarial samples can be introduced during training to prevent the model from accurately predicting disease type. Specifically, a gradient penalty term can be added during training to limit the model's dependence on high-correlation fields. For example, during training, the gradient of the model's output is calculated, and the part of the gradient related to high-correlation fields is penalized. This penalty mechanism forces the model to pay more attention to other features, reducing its dependence on high-correlation fields. In this way, the trained privacy protection defense model can effectively resist inference attacks and protect user privacy. The collaborative interaction of the federal adversarial learning network and the differential privacy de-correlation processing flow further enhances privacy protection capabilities. For example, in the federal learning framework, an adversarial learning network can generate adversarial samples to interfere with attackers' reasoning about the model. At the same time, differential privacy technology reduces the correlation between data by adding random noise to the data. For example, during the aggregation of model parameters, random noise can be added to the parameters to prevent attackers from inferring specific participants' data information from the parameters. This collaborative processing flow not only improves the privacy protection capabilities of the model but also optimizes the data partitioning and storage mechanism.For example, in a distributed storage system, data can be divided into multiple parts and stored in different participants. Through differential privacy technology, the data of each participant cannot be used to infer complete information alone, while the collaborative use of data is realized through the federated learning framework. This mechanism not only guarantees data privacy, but also improves the efficiency of data use.

[0027] In step S103, the high-sensitive field is subjected to differential privacy perturbation processing, and the correlation risk index of the high-correlation field combination is calculated in real time to generate a correlation risk analysis report.

[0028] When performing differential privacy perturbation on high-sensitivity fields, it is necessary to first determine the privacy protection requirements of sensitive fields. For example, identity card numbers and mobile phone numbers are directly linked to personal identity information, and the risk of leakage is extremely high. In order to protect these data, differential privacy technology is used to mask the true value by adding random noise to the data. In specific implementation, Laplace mechanism is used to generate noise, and the amount of noise added is adjusted according to the sensitivity of the field. For example, the sensitivity of the identity card number is high, so the range of noise added is large, while the noise range of the mobile phone number is relatively small because some information may be public. This perturbation process makes it impossible for attackers to accurately infer the original value even if they obtain the perturbed data, effectively protecting personal privacy. The core of differential privacy is to balance data usability and privacy protection, and by reasonably setting the privacy budget parameter ε, the privacy leakage risk can be minimized while ensuring the accuracy of data statistical analysis. When calculating the correlation risk index of high-correlation field combinations in real time, it is necessary to first determine which fields have high correlation. For example, identity card numbers and addresses may have strong correlation in the data set because addresses are usually bound to identity card numbers. By calculating the Pearson correlation coefficient, the correlation can be quantified. For example, the calculation result shows that the correlation coefficient of identity card number and address is 0.85, which is much higher than the preset threshold of 0.7, so these two fields are identified as high-correlation combinations. Next, based on the principle of information entropy, the risk degree of privacy leakage caused by this correlation is calculated. Information entropy reflects the uncertainty of data, the higher the correlation, the lower the information entropy, and the higher the privacy leakage risk. By real-time calculation, the privacy protection strategy can be dynamically adjusted, such as further perturbation of the address field to reduce the correlation risk. When generating the correlation risk analysis report, the above calculation results are systematically organized. The report includes the identification results of high-correlation field combinations, the quantified value of correlation risk index, and the corresponding privacy protection suggestions. For example, the report will clearly indicate that the correlation risk index of identity card number and address is 0.92, which belongs to the high-risk level, and it is recommended to perform differential privacy processing on the address field and restrict the access permission of this field. By generating such a report, data managers can clearly understand the privacy risk distribution in the data and take targeted protection measures to improve the level of data security management.

[0029] Step S104, dynamically adjust the data segmentation strategy according to the correlation risk analysis report, and optimize the selection of local model training samples; if the correlation risk index exceeds the preset safety threshold, trigger the adaptive update of the differential privacy de-correlation processing flow.

[0030] According to the correlation risk analysis report, the data segmentation strategy is dynamically adjusted, and the local model training sample selection is optimized. First of all, the core content of the correlation risk analysis report needs to be clear. The report usually contains the correlation strength between data fields, potential privacy leakage risk points, and the risk level of different data combinations. For example, in medical data, the patient's age, gender, and disease type may form a high correlation combination, which is easy to leak individual privacy through reasoning. According to the report, the data segmentation strategy can be dynamically adjusted, such as splitting the high correlation field combination into multiple low correlation subsets. For example, store age and disease type in different nodes, and when training the local model, prefer to select data with low correlation fields for training to avoid the model relying too much on high correlation fields, thereby reducing the risk of privacy leakage. When the correlation risk index exceeds the preset safety threshold, trigger the adaptive update of the differential privacy de-correlation processing procedure. Differential privacy technology protects individual privacy by adding noise to the data, but the selection of noise parameters needs to be dynamically adjusted according to the specific correlation risk index. For example, in user behavior data analysis, if it is found that the correlation risk index of some behavior patterns is high, the optimal noise parameter combination can be dynamically searched through the immune optimization algorithm. Specifically, the immune optimization algorithm simulates the diversity mechanism of the biological immune system to select the optimal solution that can effectively reduce the correlation risk without excessively affecting the data usability from multiple noise parameter combinations. For example, select the noise parameter to control the error of the data aggregation result within 5%, and reduce the correlation risk index below the safety threshold. When optimizing the local model training sample selection strategy, the transfer learning method can play an important role. Transfer learning uses knowledge from other fields or tasks to improve the performance of the current task. For example, in a financial risk control model, if the correlation risk of some fields is high, relevant feature knowledge can be transferred from low-risk fields to optimize the selection of local training samples. Specifically, samples with low correlation to high-risk fields can be selected for training, and the transfer learning model can be used to indirectly learn the high-risk fields to avoid direct exposure of sensitive information. For example, through transfer learning, the user's credit risk can be indirectly predicted using consumption behavior data without directly using the user's income field, thereby protecting privacy while improving model accuracy. For high correlation field combinations, the Apriori algorithm is used for intelligent segmentation, and the minimum support and minimum confidence requirements need to be set. For example, in e-commerce data, user purchase records and browsing records may form a high correlation combination. Through the Apriori algorithm, frequent field combinations can be identified, and according to the minimum support and minimum confidence requirements, they can be divided into multiple low correlation subsets. For example, store purchase records and browsing records in different nodes, and build a distributed hash table index to ensure data query efficiency.Distributed hash table indexing can effectively reduce the risk of data leakage from a single node by storing data across different nodes, while improving the concurrency performance of data access. For example, in a distributed storage system, each node only stores part of the data, and when querying, the target node is quickly located through the hash table, thereby protecting privacy while improving system performance.

[0031] In step S105, the federated adversarial learning network and the differential privacy de-association processing flow are cooperatively interacted to continuously optimize the data partition storage mechanism.

[0032] The federated adversarial learning network trains the model locally through multiple participants and uploads the model parameters to the central server for aggregation, avoiding direct transmission of raw data. In this process, the differential privacy de-association processing flow is introduced, which reduces the risk of data leakage by adding noise to the model parameters. For example, in the medical data sharing scenario, electronic medical record data from different hospitals is jointly modeled through the federated learning framework, but when uploading the model parameters, Laplace noise is added to ensure that the specific patient's sensitive information cannot be inferred from the model parameters. This cooperative interaction mechanism not only protects data privacy, but also ensures the training effect of the model. The data partition storage mechanism is another key link in privacy protection. By dividing data into multiple segments and storing them on different servers or nodes, even if a node is attacked, the attacker cannot obtain the complete raw data. For example, in the financial field, user transaction data is divided into multiple parts and stored on different cloud servers. Only through the encrypted aggregation of the federated learning framework can the global model be updated. This mechanism further reduces the risk of data leakage while improving the fault tolerance of the system. When the correlation risk index exceeds the preset security threshold, the differential privacy de-association processing flow triggers adaptive updates. Through the immune optimization algorithm, the optimal noise parameter combination is dynamically searched to ensure that privacy is protected while minimizing the impact on model accuracy. For example, in an intelligent recommendation system, user behavior data is modeled through federated learning, and when the system detects that some user behavior data may expose their identity, it automatically adjusts the noise parameters to meet privacy protection requirements while not affecting the accuracy of the recommendation results. In the local initialization of the CNN convolutional neural network model parameters by the participants, the federated learning framework FederatedAveraging uses encrypted aggregation and asynchronous updating to ensure the security and consistency of the global model. For example, in the field of autonomous driving, sensor data from different vehicles is trained through local CNN models and uploaded to the central server for encrypted aggregation. Due to the use of asynchronous updating mechanism, even if some vehicles have large communication delays, it will not affect the training progress of the global model. This design not only improves the robustness of the system, but also ensures the security of data privacy.

[0033] Step S106, the original data set is obtained and the high correlation field combination is determined, comprising: extracting the identity card number and mobile phone number from the original data set as high sensitive fields; calculating the Pearson correlation coefficient matrix between the high sensitive fields and other fields; according to the Pearson correlation coefficient matrix, determining the high correlation field combination with the correlation coefficient greater than the preset threshold.

[0034] Extracting the identity card number and mobile phone number from the original data set as high sensitive fields is because these information has high personal privacy and uniqueness, and once leaked may cause serious privacy risk. The identity card number usually contains the personal date of birth, gender and region information, and the mobile phone number is an important identification for personal communication and identity verification. Therefore, it is reasonable to define these fields as high sensitive fields. In actual operation, these fields can be quickly extracted from the original data through data preprocessing techniques such as regular expressions or field matching. Calculating the Pearson correlation coefficient matrix between the high sensitive fields and other fields is to quantify their linear relationship. The Pearson correlation coefficient ranges from negative one to positive one, and the closer the value is to positive one or negative one, the stronger the linear relationship between the two fields. For example, suppose in a certain data set, the correlation coefficient between the identity card number and the date of birth is 0.98, indicating a high correlation between the two. While the correlation coefficient between the mobile phone number and the residence address is 0.12, indicating that the two are almost unrelated. By constructing the correlation coefficient matrix, the correlation degree of high sensitive fields and other fields can be intuitively displayed. According to the Pearson correlation coefficient matrix, the high correlation field combination with the correlation coefficient greater than the preset threshold is determined, in order to identify the field combination that poses a potential threat to privacy protection. The preset threshold can be set according to specific business needs, such as 0.7 or 0.8. When the correlation coefficient exceeds the threshold, it means that there is a strong correlation between the two fields, which may be used for inference attacks. For example, suppose in a certain data set, the correlation coefficient between the identity card number and the name is 0.85, which exceeds the preset threshold, then these two fields are considered as a high correlation field combination. Such combination needs to be paid attention to, because the attacker may infer the identity card number from the name, or vice versa. Through the above steps, the correlation between high sensitive fields and other fields can be systematically identified, providing data support for subsequent privacy protection measures. Calculating the Pearson correlation coefficient matrix not only quantifies the relationship between fields, but also helps to screen out field combinations that pose a potential threat to privacy. The advantage of this method is its efficiency and interpretability, which can provide a scientific basis for the design of privacy protection models.

[0035] In step S107, the intelligent segmentation and dispersed storage of the high-correlation field combination comprises: using an Apriori algorithm to analyze the high-correlation field combination; intelligently segmenting the high-correlation field combination according to minimum support and minimum confidence requirements; dispersively storing the segmented data into different data nodes; and constructing a distributed hash table index based on the dispersedly stored data.

[0036] To analyze high-correlation field combinations using the Apriori algorithm, first understand the core idea of the Apriori algorithm. The Apriori algorithm is a classic algorithm for mining frequent itemsets and association rules, which finds frequent itemsets that meet the minimum support from the dataset through layer-by-layer search. In this scenario, high-correlation field combinations can be considered as frequent itemsets, and the algorithm will generate candidate itemsets through multiple scans of the dataset and filter out combinations with support exceeding the preset threshold. For example, assuming there is a dataset containing fields such as ID number, mobile number, and address, the algorithm will first calculate the support of each field, then generate the support of two-by-two combinations, and finally determine which combinations are highly correlated. According to the minimum support and minimum confidence requirements, intelligent segmentation of high-correlation field combinations is performed, which is the key to ensuring data dispersion and privacy protection. The minimum support refers to the minimum frequency of a field combination appearing in the entire dataset, while the minimum confidence refers to the predictive ability of one field appearing in another field. By setting these two thresholds, we can filter out field combinations that are truly highly correlated. For example, assuming the minimum support is set to 0.1 and the minimum confidence is set to 0.8, the algorithm will filter out combinations that appear more than 10% of the time in the dataset and have a correlation between fields exceeding 80%, and then segment these combinations to ensure that sensitive information is not stored centrally. Store the segmented data in different data nodes to achieve distributed storage and enhance system scalability. By segmenting high-correlation field combinations into multiple parts and storing them in different data nodes, we can effectively reduce the data load on individual nodes and improve system processing efficiency. For example, if the combination of ID number and mobile number is segmented into two parts, with ID number stored in node A and mobile number stored in node B, even if a node is attacked, the attacker cannot obtain complete sensitive information, thereby enhancing data security. Build a distributed hash table index based on the dispersedly stored data to achieve fast query and efficient management of data. The distributed hash table is an index structure used to quickly find data in a distributed system, which maps data to different nodes through a hash function and maintains a global index table. For example, assuming that the segmented data of ID number and mobile number is stored in nodes A and B respectively, the distributed hash table will record the storage location of each field, and when a field needs to be queried, the system can quickly locate the corresponding node, thereby improving data query efficiency. Through the above steps, intelligent segmentation and dispersed storage of high-correlation field combinations can be achieved, protecting the privacy of sensitive data while improving system processing efficiency and scalability.The core of the process is to screen out high correlation field combinations through the Apriori algorithm, and to realize efficient management of data through distributed storage and hash table index, so as to achieve the dual goals of data security and system performance.

[0037] Step S108, the encryption aggregation and asynchronous update of the global model parameter is realized through the federated learning framework, comprising: initializing the CNN convolutional neural network model parameter locally at the participant; realizing the encryption aggregation of the CNN convolutional neural network model parameter through the FederatedAveraging framework; and asynchronously updating the CNN convolutional neural network model parameter.

[0038] When each participant initializes the CNN convolutional neural network model parameters locally, each participant independently constructs the model structure and initializes the parameters according to its own local data distribution. For example, participant A and participant B each have different image data sets, A's data set is mainly composed of cats and dogs, and B's data set is mainly composed of birds. Both parties use the same CNN architecture, but use their own random seeds when initializing the parameters to ensure that the initial values of the parameters are different. This local initialization method can preserve the distribution characteristics of the data and avoid the model from being biased towards the data distribution of one party at the beginning of training, thereby laying the foundation for subsequent federated learning. When implementing encrypted aggregation of CNN convolutional neural network model parameters through the FederatedAveraging framework, each participant uploads the model parameters to the central server after local training. For example, participants A and B each train for 10 rounds and send their respective model weight matrices (such as convolution kernel parameters and fully connected layer weights) to the server. The server uses secure multi-party computation technology to encrypt and aggregate the parameters, ensuring that the original parameters of individual participants cannot be directly obtained during the aggregation process. Specifically, the server performs weighted averaging on the received encrypted parameters, for example, the weights of A and B are 0.6 and 0.4 respectively, and the final global model parameters are the weighted sum of the two. This encrypted aggregation method protects data privacy while optimizing the global model. When updating the CNN convolutional neural network model parameters asynchronously, each participant updates the local model according to its own computing resources and data characteristics after receiving the global model parameters. For example, participant A has sufficient computing resources and can immediately perform 5 rounds of local training after receiving the global parameters; while participant B has limited computing resources and can only perform 2 rounds of local training. This asynchronous update mechanism allows different participants to adjust the training pace flexibly according to their own conditions, avoiding delays caused by one party affecting the overall training efficiency. At the same time, asynchronous updating can effectively deal with network fluctuations and participant dropouts, ensuring the robustness of the federated learning system. Through the above steps, the federated learning framework can efficiently train and optimize the global model while protecting data privacy. Local initialization ensures model diversity, encrypted aggregation protects data privacy, and asynchronous updating improves system flexibility and robustness. The synergy of these technologies enables federated learning to effectively address data silos and privacy protection issues in practical applications, providing reliable technical support for multi-participant collaboration.

[0039] In step S109, the high-correlation field combination is used to construct an adversarial sample and train a privacy protection defense model. The adversarial sample is constructed for Member inference attacks based on the high-correlation field combination. A gradient penalty term is introduced during the training process. The privacy protection defense model is trained using the adversarial sample and the gradient penalty term.

[0040] To construct adversarial samples and train privacy protection defense models for high correlation field combinations, the definition of high correlation field combinations and their risks in privacy leakage need to be clarified first. High correlation field combinations usually refer to fields with strong statistical correlation in the dataset, such as ID number and mobile phone number, name and address, etc. Once these field combinations are exploited by attackers, sensitive information may be restored through inference attacks. Therefore, the construction of adversarial samples for Member inference attacks aims to generate data samples that can confuse or interfere with the reasoning ability of attackers by simulating the behavior of attackers. For example, when constructing adversarial samples, a small perturbation can be made to high correlation fields, such as randomly replacing some digits of a mobile phone number, thereby reducing its correlation with the ID number. Such perturbation not only preserves the overall distribution characteristics of the data, but also effectively prevents attackers from restoring sensitive information through correlation reasoning. The introduction of a gradient penalty term during training is to further optimize the robustness of the privacy protection defense model. The core idea of the gradient penalty term is to constrain the gradient change of the model during training to prevent the model from overfitting to the high correlation features in the training data, thereby reducing the vulnerability of the model in inference attacks. For example, during training, the gradient of the model loss function can be calculated and a certain penalty weight can be applied to it, so that the model not only focuses on prediction accuracy during optimization, but also focuses on gradient smoothness. This method can effectively prevent the model from experiencing a significant performance decline under adversarial sample attacks, while improving the model's generalization ability for high correlation field combinations. By training privacy protection defense models with adversarial samples and gradient penalty terms, the ultimate goal is to build a model that can effectively resist Member inference attacks in real-world scenarios. During training, adversarial samples simulate the behavior of attackers, while the gradient penalty term is used to constrain the optimization direction of the model. For example, during training, adversarial samples can be mixed with original data input into the model, while the gradient penalty term is introduced as an additional loss function. This method allows the model to learn how to maintain performance stability under adversarial sample attacks during training, and prevents the model from relying too much on high correlation field combinations through the gradient penalty term. Ultimately, the trained model not only performs well on regular data, but also exhibits stronger robustness and privacy protection capabilities when facing Member inference attacks.

[0041] In step S1010, the high-sensitive field is subjected to differential privacy perturbation processing, including: adopting a maximum mean difference method to perform feature transformation on the original data; generating differential privacy random noise based on a Laplace mechanism; and using the differential privacy random noise to perturb the high-sensitive field.

[0042] The maximum mean difference method is a feature transformation technique designed to extract significant features by calculating the mean difference of data, thereby reducing the sensitivity of the data. In implementation, first, the mean value of high-sensitive fields (such as ID number, mobile phone number) in the original data is calculated, and then the mean difference of these fields and other fields is compared to determine their significant features. For example, for the mobile phone number field, the mean value of its digits can be calculated and compared with the mean values of other fields such as age and gender to extract key information that reflects user characteristics. This transformation can effectively reduce the sensitivity of the data while retaining useful information, laying the foundation for subsequent differential privacy processing. Differential privacy random noise generation based on the Laplace mechanism is a classic privacy protection technique, whose core idea is to add random noise conforming to the Laplace distribution to mask sensitive data. The characteristic of the Laplace distribution is that it can adjust the strength of the noise according to the sensitivity of the data, thereby protecting privacy while minimizing the impact on data accuracy. For example, for the mobile phone number field, assuming its sensitivity is 10, the strength of the generated Laplace noise will be adjusted according to this sensitivity to ensure that the added noise can protect privacy and not significantly affect the overall distribution of the data. This mechanism can effectively prevent attackers from inferring sensitive information about users through data analysis. Using differential privacy random noise to perturb high-sensitive fields is the core step of differential privacy protection. In implementation, the generated Laplace noise is added to the high-sensitive fields to perturb the data. For example, for the mobile phone number field, assuming the original value is "12345678910" and the generated noise is "3", the perturbed value is "12345678913". This perturbation can effectively prevent attackers from obtaining users' real information through data matching or reasoning, while ensuring that the statistical properties of the data remain usable. In this way, user privacy can be protected while supporting the normal performance of data analysis tasks. The technical effects of the above methods mainly include the following aspects: first, the maximum mean difference method reduces the sensitivity of the data through feature transformation, providing a foundation for subsequent differential privacy processing; second, the Laplace mechanism can generate appropriate noise according to the sensitivity of the data, ensuring the effectiveness of privacy protection; finally, using noise to perturb high-sensitive fields can protect privacy while retaining data usability. These methods together form a rigorous privacy protection system that can effectively address the risks of data leakage and privacy infringement.

[0043] In step S1011, the real-time calculation of the correlation risk index of the high-correlation field combination and the generation of the correlation risk analysis report include: calculating the correlation risk index of the high-correlation field combination based on the information entropy principle; generating a data correlation risk analysis report according to the correlation risk index; and taking the data correlation risk analysis report as the basis for subsequent dynamic adjustment.

[0044] To calculate the correlation risk index of high-correlation field combinations based on the information entropy principle, first, the fields in the data nodes need to be analyzed to identify which fields have high correlation. Information entropy is an indicator of data uncertainty. By calculating the information entropy of a field, the dependency between fields can be evaluated. For example, in a dataset containing "age" and "income" fields, if the information entropy of the "age" field is low, it indicates that the values of this field are concentrated, while the information entropy of the "income" field is high, indicating that its values are more dispersed. By analyzing the joint information entropy of these two fields, the correlation between them can be determined. If the joint information entropy is significantly lower than the sum of the individual information entropies, it indicates that there is high correlation between the two fields. Further, the correlation risk index can be calculated to quantify the risk that this high correlation may bring. For example, if the correlation risk index of the "age" and "income" fields is high, it indicates that the combination of these two fields may leak sensitive information, such as a specific age group with high income. According to the correlation risk index, a data correlation risk analysis report is generated, which compares the calculated correlation risk index with the preset security threshold and generates the corresponding analysis results. For example, if the correlation risk index of the "age" and "income" fields is 0.85, and the preset security threshold is 0.7, this field combination is considered to have a high correlation risk. In the report, these high-correlation field combinations will be listed in detail, along with their correlation risk indexes. In addition, the report can include detailed information such as field distribution and correlation strength for subsequent analysis. For example, the report can indicate that the "age" field values are concentrated between 30-40 years old, while the "income" field values are between 5000-10000 yuan, and the correlation strength between these two fields is 0.9, indicating a strong dependency between them. The data correlation risk analysis report serves as the basis for subsequent dynamic adjustments, allowing for adjustments to data segmentation granularity and dimension based on the analysis results in the report. For example, if the correlation risk index of the "age" and "income" fields is high, the data can be segmented more finely according to the "age" field, or a new dimension such as "occupation" can be added to reduce the correlation between fields. After adjustment, the distribution balance index of the fields needs to be recalculated to ensure that the segmented data is balanced in each dimension. For example, if the data is segmented into two subsets of 30-35 years old and 35-40 years old according to the "age" field, it needs to be checked whether the distribution of the "income" field in these two subsets is balanced. If the "income" field values in a certain subset are too concentrated, the segmentation strategy needs to be further adjusted. In this way, the data segmentation scheme can be dynamically optimized to reduce the risk brought by high-correlation field combinations.

[0045] In step S1012, the data segmentation strategy is dynamically adjusted and the local model training sample selection is optimized according to the correlation risk analysis report, including: adjusting the data segmentation granularity and dimension according to the correlation risk analysis report; recalculating the field distribution balance index; and optimizing the local model training sample selection strategy by a transfer learning method.

[0046] To dynamically adjust the data segmentation strategy and optimize the local model training sample selection according to the correlation risk analysis report, first, the granularity and dimension of data segmentation are adjusted according to the data correlation evaluation result in the correlation risk analysis report. For example, in a certain financial scenario, if it is found that the correlation risk between user ID and transaction amount is high, the data segmentation granularity can be adjusted from the original user ID segmentation to transaction time segmentation, and the transaction amount segmentation dimension is added to reduce the data correlation. In specific implementation, the most suitable strategy can be selected by analyzing the data distribution characteristics under different segmentation granularity and dimension. For example, after the data is segmented by day, the field distribution balance index is recalculated, and it is found that the distribution of transaction amount is more uniform, thereby reducing the risk of data leakage. After adjusting the data segmentation strategy, the field distribution balance index needs to be recalculated to evaluate the effect of data segmentation. For example, in a certain medical data scenario, if the patient data is segmented by age, the distribution balance index of the gender field is recalculated, and it is found that the balance of gender distribution is significantly improved, indicating that the segmentation strategy effectively reduces the data correlation. In specific implementation, the distribution of fields under different segmentation strategies can be counted, the balance index can be calculated, and the optimal strategy can be selected. For example, by calculating the variance or standard deviation of a certain field under different segmentation strategies, the distribution balance is evaluated, and the strategy with the smallest variance is selected as the final solution. By optimizing the local model training sample selection strategy through a transfer learning method, the model performance can be further improved. For example, in a certain image classification task, if the local data sample is small, features can be extracted from a pre-trained model in other related fields through transfer learning to optimize the training sample selection of the local model. In specific implementation, features can be first extracted from the pre-trained model, and then the most relevant samples to the local task are selected for training according to the feature similarity. For example, by calculating the cosine similarity between local samples and pre-trained samples, the samples with the highest similarity are selected as training data, thereby improving the classification accuracy of the model. The implementation of the above steps not only reduces the data correlation risk, but also improves the balance of data segmentation and the efficiency of model training. For example, in an e-commerce scenario, by adjusting the data segmentation strategy and optimizing the sample selection strategy, it is found that the accuracy of the product recommendation model is improved by 5%, and the data leakage risk is reduced by 10%. This technical effect has important significance in practical application, which can effectively protect user privacy and improve model performance.

[0047] The above examples are only used to illustrate the technical solutions of the present application but not to limit the present application, and the present application is described in detail only with reference to the preferred embodiments. It should be understood by those of ordinary skill in the art that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A data privacy protection method, characterized in that, include: Obtain the original dataset, extract highly sensitive fields and calculate their correlation with other fields to determine the combination of highly correlated fields; The highly correlated field combinations are intelligently segmented and stored in a distributed manner to construct a distributed hash table index; the intelligent segmentation involves dividing the highly correlated field combinations into multiple subsets according to preset rules. A federated learning framework is used to achieve encrypted aggregation and asynchronous updating of global model parameters; Adversarial examples are constructed for the highly correlated field combinations to train a privacy-preserving defense model; Differential privacy perturbation processing is applied to the highly sensitive fields; The association risk index of the highly correlated field combination is calculated in real time, and an association risk analysis report is generated. Based on the aforementioned risk analysis report, the data segmentation strategy is dynamically adjusted to optimize the selection of local model training samples; If the association risk index exceeds the preset security threshold, an adaptive update of the differential privacy deassociation process is triggered. Through the collaborative interaction of federated adversarial learning networks and differential privacy deassociation processes, the data segmentation and storage mechanism is continuously optimized.

2. The method as described in claim 1, characterized in that, The process of obtaining the original dataset and determining the highly relevant field combinations includes: Extract the ID card number and mobile phone number as highly sensitive fields from the original dataset; Calculate the Pearson correlation coefficient matrix between the highly sensitive field and other fields; Based on the Pearson correlation coefficient matrix, determine the high correlation field combinations with correlation coefficients greater than a preset threshold.

3. The method as described in claim 1, characterized in that, Intelligent segmentation and distributed storage of the highly correlated field combinations include: The Apriori algorithm was used to analyze the highly correlated field combinations; Intelligent segmentation is performed on the highly correlated field combinations based on minimum support and minimum confidence requirements; The split data is distributed and stored across different data nodes; A distributed hash table index is constructed based on the data stored in a dispersed manner.

4. The method as described in claim 1, characterized in that, The method of implementing encrypted aggregation and asynchronous updating of global model parameters through a federated learning framework includes: Initialize the CNN convolutional neural network model parameters locally on the participating party; The encrypted aggregation of parameters for the CNN convolutional neural network model is achieved using the FederatedAveraging framework. The parameters of the CNN convolutional neural network model are updated asynchronously.

5. The method as described in claim 1, characterized in that, Construct adversarial examples for the highly correlated field combinations and train a privacy-preserving defense model, including: Based on the combination of the highly correlated fields, construct adversarial samples for Member inference attacks; Introduce a gradient penalty term during training; The privacy-preserving defense model is trained using the adversarial examples and gradient penalty terms.

6. The method as described in claim 1, characterized in that, Differential privacy perturbation processing is applied to the highly sensitive fields, including: The original data are transformed using the maximum mean difference method; Differential privacy-preserving random noise is generated based on the Laplace mechanism; The highly sensitive field is perturbed using the differential privacy random noise.

7. The method as described in claim 1, characterized in that, The real-time calculation of the association risk index of highly correlated field combinations and the generation of an association risk analysis report include: The association risk index of the highly correlated field combination is calculated based on the principle of information entropy. A data-related risk analysis report is generated based on the aforementioned associated risk index; The aforementioned data-related risk analysis report will serve as the basis for subsequent dynamic adjustments.

8. The method as described in claim 1, characterized in that, Based on the aforementioned risk analysis report, the data segmentation strategy is dynamically adjusted to optimize the selection of local model training samples, including: Adjust the data segmentation granularity and segmentation dimensions based on the aforementioned associated risk analysis report; Recalculate the field distribution balance index; The local model training sample selection strategy is optimized through transfer learning.

Images