Data privacy protection method

Abstract

The invention provides a data privacy protection method, which comprises the following steps of: acquiring an original data set, extracting a high-sensitivity field, calculating the correlation between the high-sensitivity field and other fields, and determining a high-correlation field combination; performing intelligent segmentation and scattered storage on the high-relevance field combination, and constructing a distributed hash table index; realizing encryption aggregation and asynchronous updating of global model parameters through a federated learning framework; constructing an adversarial sample for the high-relevance field combination, and training a privacy protection defense model; differential privacy perturbation processing is carried out on the high-sensitivity field; calculating an association risk index of the high association field combination in real time, and generating an association risk analysis report; dynamically adjusting a data segmentation strategy according to the associated risk analysis report, and optimizing local model training sample selection; and if the association risk index exceeds a preset security threshold, triggering adaptive updating of a differential privacy deassociation processing flow.

Description

Technical Field

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

[0002] Background: In business scenarios, highly sensitive fields such as ID card number and mobile phone number have complex correlation characteristics with other fields. By calculating the Pearson correlation coefficient matrix, it is possible to determine whether the correlation coefficient between these fields exceeds the preset threshold, thereby identifying highly correlated field combinations. In order to solve the privacy leakage risk brought by highly correlated field combinations, it is necessary to adopt the Apriori algorithm based on association rule mining 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 in a decentralized manner, and a distributed hash table index is constructed to form a distributed data storage structure for privacy protection. At the same time, in order to protect privacy during model training, it is necessary to build a federated adversarial learning network, and realize the encrypted aggregation and asynchronous update of global model parameters through the federated learning framework. In the privacy protection of highly correlated field combinations, it is also necessary to construct adversarial samples for Member reasoning attacks, and train the privacy protection defense model by introducing gradient penalty terms, so that it can identify and resist privacy reasoning attack behaviors based on data correlation. In order to further reduce the correlation between fields, the maximum mean difference method can be used to transform the features of the original data, and the Laplace mechanism can be used to generate differential privacy random noise to perturb the fields with higher sensitivity. By establishing a dynamic assessment system for association risk, the association risk index of highly correlated field combinations in data nodes is calculated in real time based on the principle of information entropy, and a data association risk analysis report can be generated to quantify the risk level of privacy leakage. According to the association risk analysis report, the data segmentation granularity and segmentation 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 security 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 data availability loss. Finally, through the collaborative 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 storage mechanism is continuously optimized, the risk of leakage of sensitive privacy information is minimized, and data security sharing is achieved while ensuring the value of data analysis. Summary of the invention

[0003] The present invention provides data security management for federated learning, mainly including: Obtain the original data set, extract the highly sensitive fields and calculate the correlation between them and other fields, and determine the highly correlated field combination; intelligently segment and disperse the storage for the highly correlated field combination, and construct a distributed hash table index; realize the encrypted aggregation and asynchronous update of global model parameters through the federated learning framework; construct adversarial samples for the highly correlated field combination, and train the privacy protection defense model; perform differential privacy perturbation processing on the highly sensitive fields; calculate the association risk index of the highly correlated field combination in real time, and generate an association risk analysis report; dynamically adjust the data segmentation strategy according to the association risk analysis report, and optimize the selection of local model training samples; if the association risk index exceeds the preset security threshold, trigger the adaptive update of the differential privacy de-correlation processing flow; through the collaborative interaction of the federated adversarial learning network and the differential privacy de-correlation processing flow, continuously optimize the data segmentation storage mechanism. Further, the acquisition of the original data set and determination of the highly correlated field combination include: extracting the ID card number and the mobile phone number from the original data set as highly sensitive fields; calculating the Pearson correlation coefficient matrix between the highly sensitive fields and other fields; according to the Pearson correlation coefficient matrix, determining the highly correlated field combination whose correlation coefficient is greater than the preset threshold. Furthermore, the intelligent segmentation and decentralized storage of highly correlated field combinations include: analyzing the highly correlated field combinations using the Apriori algorithm; intelligently segmenting the highly correlated field combinations according to the minimum support and minimum confidence requirements; distributing and storing the segmented data in different data nodes; and constructing a distributed hash table index based on the decentralized stored data. Furthermore, the encrypted aggregation and asynchronous update of global model parameters are achieved through a federated learning framework, including: initializing CNN convolutional neural network model parameters locally on the participating parties; implementing encrypted aggregation of the CNN convolutional neural network model parameters through a FederatedAveraging framework; and asynchronously updating the CNN convolutional neural network model parameters. Furthermore, the construction of adversarial samples for highly correlated field combinations and training of privacy protection defense models include: constructing adversarial samples for Member reasoning attacks based on the highly correlated field combinations; introducing gradient penalty terms during the training process; and training the privacy protection defense model through the adversarial samples and gradient penalty terms. Furthermore, the differential privacy perturbation processing of the highly sensitive field includes: using the maximum mean difference method to transform the features of the original data; generating differential privacy random noise based on the Laplace mechanism; and using the differential privacy random noise to perturb the highly sensitive field.Furthermore, the real-time calculation of the correlation risk index of the high correlation field combination and the generation of the correlation risk analysis report includes: 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 using the data correlation risk analysis report as a basis for subsequent dynamic adjustments. Furthermore, the dynamic adjustment of the data segmentation strategy and optimization of the local model training sample selection according to the correlation risk analysis report includes: adjusting the data segmentation granularity and 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.

[0004] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses a privacy-preserving data processing method based on federated learning. The method first extracts highly sensitive fields from the original data set, analyzes their correlation with other fields, and determines highly correlated field combinations. Then, these field combinations are intelligently segmented and dispersedly stored to construct a distributed hash table index. The encrypted aggregation and asynchronous update of model parameters are realized through a federated learning framework, and a privacy-preserving defense model for adversarial sample training is constructed at the same time. Differential privacy perturbation processing is performed on highly sensitive fields, and the associated 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 the threshold, the adaptive update of the differential privacy de-correlation processing is triggered. The present invention continuously optimizes the data segmentation storage mechanism through the collaboration of federated adversarial learning and differential privacy, effectively protects the privacy of sensitive data, and ensures data availability. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] Figure 1 The flowchart of data security management for federated learning of the present invention.

[0006] Figure 2 A schematic diagram of data security management for federated learning according to the present invention.

[0007] Figure 3 This is another schematic diagram of data security management for federated learning according to the present invention.

[0008] Figure 4 This is another schematic diagram of data security management for federated learning according to the present invention.

[0009] Figure 5 This is another schematic diagram of data security management for federated learning according to the present invention.

[0010] Figure 6 This is another schematic diagram of data security management for federated learning according to the present invention.

[0011] Figure 7 This is another schematic diagram of data security management for federated learning according to the present invention.

[0012] Figure 8 This is another schematic diagram of data security management for federated learning according to the present invention.

[0013] Fig. 9 This is another schematic diagram of data security management for federated learning according to the present invention.

[0014] Fig.10 This is another schematic diagram of data security management for federated learning according to the present invention.

[0015] Fig.11 This is another schematic diagram of data security management for federated learning according to the present invention.

[0016] Fig.12 This is another schematic diagram of data security management for federated learning according to the present invention.

[0017] Fig.13 This is another schematic diagram of data security management for federated learning according to the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] like Figure 1-13 The data security management for federated learning in this embodiment may specifically include: Step S101, obtaining the original data set, extracting highly sensitive fields and calculating the correlation between them and other fields, and determining a highly correlated field combination; performing intelligent segmentation and distributed storage on the highly correlated field combination, and constructing a distributed hash table index.

[0020] When obtaining the original data set, you first need to clarify the data source and structure. Assume that the data set comes from a customer information table of a financial company, which contains fields such as ID number, mobile phone number, name, address, and consumption record. Among them, ID number and mobile phone number are highly sensitive fields and need to be extracted first. After extracting the highly sensitive fields, the linear relationship between them can be quantified by calculating the Pearson correlation coefficient matrix between them and other fields. For example, the correlation coefficient between the ID number and the 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 highly correlated field combination, such as ID number and address, mobile phone number and consumption record, etc. When performing intelligent segmentation on highly correlated field combinations, the Apriori algorithm can be used. The algorithm finds frequent item sets and generates association rules by setting the minimum support and minimum confidence. For example, taking the ID number and address as a frequent item set, setting the minimum support to 0.5 and the minimum confidence to 0.8, these two fields will only be retained if the frequency of co-occurrence exceeds 50% and the correlation strength exceeds 80%. Subsequently, according to these rules, the highly correlated field combinations are split into multiple subsets and stored in different data nodes. For example, the ID card number and address are stored in one node, and the mobile phone number and consumption record are stored in another node. The purpose of building a distributed hash table index is to improve the efficiency of data query. By performing hash calculations on the segmented field combinations, unique hash values ​​are generated, and these hash values ​​are mapped to the corresponding data nodes. For example, a hash value is obtained by performing a hash calculation on the combination of the ID card number and the address, and then the hash value is associated with the data node storing the combination. In this way, when querying, you only need to quickly locate the corresponding node based on the hash value, without traversing the entire data set, thereby improving query efficiency and reducing the risk of data leakage.

[0021] Step S102, realizing encrypted aggregation and asynchronous update of global model parameters through a federated learning framework; constructing adversarial samples for the highly correlated field combinations, and training a privacy protection defense model.

[0022] To achieve encrypted aggregation and asynchronous update of global model parameters through the federated learning framework, we first need to understand the basic principles of federated learning. Federated learning is a distributed machine learning method in which participants train models locally instead of centralizing data to a central server. To ensure data privacy, the parameters of the global model are aggregated through encryption technology after being updated locally. For example, suppose there are three participants A, B, and C. Each participant trains a CNN model locally and generates model parameters W_A, W_B, and W_C respectively. These parameters are uploaded to the central server through encryption technology (such as homomorphic encryption or secure multi-party computing). The server performs weighted average of the parameters without decryption to obtain the global model parameter W_global. This encrypted aggregation method ensures the data privacy of the participants and realizes the collaborative training of the model. Asynchronous update allows participants to upload model parameters without complete synchronization, which improves the flexibility and efficiency of the system. 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 the computing power or network conditions of the participants are inconsistent. To construct adversarial samples for highly correlated field combinations and train privacy-preserving defense models, it is first necessary to identify highly correlated fields in the data. For example, in a medical data set, the age, gender, and disease type of a patient may be highly correlated, and an attacker may infer the identity information of a specific patient through these fields. To defend against such inference attacks, adversarial samples can be generated to disrupt the output of the model. For example, assuming that a model predicts the type of disease by age and gender, adversarial samples can be introduced into the training data so that the model cannot accurately predict the type of disease. Specifically, a gradient penalty term can be added during the training process to limit the model's reliance on highly correlated fields. For example, during the training process, the gradient of the model output is calculated, and the part of the gradient related to the highly correlated field is penalized. This penalty mechanism forces the model to pay more attention to other features, thereby reducing its reliance on highly correlated fields. In this way, the trained privacy-preserving defense model can effectively resist inference attacks and protect user privacy. The collaborative interaction between the federated adversarial learning network and the differential privacy de-correlation processing flow further enhances the privacy protection capability. For example, in the federated learning framework, adversarial learning networks 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, in the process of model parameter aggregation, random noise can be added to the parameters so that attackers cannot infer the data information of specific participants through the parameters. This collaborative processing flow not only improves the privacy protection ability of the model, but also optimizes the data segmentation and storage mechanism.For example, in a distributed storage system, data can be divided into multiple parts and stored in different participants. Differential privacy technology is used to ensure that the complete information of each participant's data cannot be inferred independently, and the collaborative use of data is achieved through the federated learning framework. This mechanism improves the efficiency of data use while ensuring data privacy.

[0023] Step S103, performing differential privacy perturbation processing on the highly sensitive fields; calculating the association risk index of the highly correlated field combination in real time, and generating an association risk analysis report.

[0024] When performing differential privacy perturbation processing on highly sensitive fields, it is first necessary to clarify the privacy protection requirements of sensitive fields. Taking the ID card number and mobile phone number as examples, these fields are directly related to personal identity information and have a very high risk of leakage. In order to protect this data, differential privacy technology is used to cover up the true value by adding random noise to the data. In the specific implementation, the 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 ID card number has a high sensitivity, so the added noise range is large, while the mobile phone number has a relatively small noise range because some information may be public. This perturbation processing makes it impossible for the attacker to accurately infer the original value even if he obtains the perturbed data, thereby effectively protecting personal privacy. The core of differential privacy is to balance data availability and privacy protection. By reasonably setting the privacy budget parameter ε, the risk of privacy leakage can be minimized while ensuring the accuracy of data statistical analysis. When calculating the correlation risk index of a combination of highly correlated fields in real time, it is first necessary to determine which fields have a high correlation. Taking the ID card number and address as an example, these two fields may have a strong correlation in the data set because the address is usually bound to the ID card number. By calculating the Pearson correlation coefficient, this correlation can be quantified. For example, the calculation results show that the correlation coefficient between the ID card number and the address is 0.85, which is much higher than the preset threshold of 0.7, so these two fields are identified as a high correlation combination. Next, based on the principle of information entropy, the risk level 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 risk of privacy leakage. Through 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 contains the identification results of the highly correlated field combination, the quantitative value of the correlation risk index, and the corresponding privacy protection recommendations. For example, the report will clearly point out that the correlation risk index between the ID card number and the address is 0.92, which belongs to the high risk level. It is recommended to perform differential privacy processing on the address field and restrict the access rights of the field. By generating such a report, data managers can clearly understand the distribution of privacy risks in the data and take targeted protection measures, thereby comprehensively improving the level of data security management.

[0025] Step S104, dynamically adjust the data segmentation strategy according to the association risk analysis report to optimize the selection of local model training samples; if the association risk index exceeds the preset security threshold, trigger the adaptive update of the differential privacy de-association processing flow.

[0026] When dynamically adjusting the data segmentation strategy based on the association risk analysis report and optimizing the selection of local model training samples, it is first necessary to clarify the core content of the association risk analysis report. The report usually contains the strength of association between data fields, potential privacy leakage risk points, and risk levels 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 personal 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, age and disease type are stored in different nodes respectively, and when training the local model, data with low-correlation fields are preferentially selected for training to avoid excessive reliance on high-correlation fields, thereby reducing the risk of privacy leakage. When the association risk index exceeds the preset security threshold, the adaptive update of the differential privacy de-correlation processing flow is triggered. 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 association risk index. For example, in the analysis of user behavior data, if it is found that the association risk index of certain 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 and selects the optimal solution from multiple noise parameter combinations that can effectively reduce the correlation risk without excessively affecting the data availability. For example, the noise parameter is selected so that the error of the data aggregation result is controlled within 5%, and the correlation risk index is reduced to below the safety threshold. Transfer learning methods can play an important role in optimizing the local model training sample selection strategy. Transfer learning improves the performance of the current task by using knowledge from other fields or tasks. For example, in the financial risk control model, if the correlation risk of some fields is high, the relevant feature knowledge can be transferred from the low-risk field to optimize the selection of local training samples. Specifically, samples with low correlation with high-risk fields can be selected for training, and the high-risk fields can be indirectly learned using the transfer learning model to avoid directly exposing sensitive information. For example, through transfer learning, the credit risk of users can be indirectly predicted using consumption behavior data without directly using the user income field, thereby improving the accuracy of the model while protecting privacy. For highly correlated field combinations, when using the Apriori algorithm for intelligent segmentation, it is necessary to set the minimum support and minimum confidence requirements. For example, in e-commerce data, a user's purchase history and browsing history may form a highly correlated combination. The Apriori algorithm can identify frequently occurring field combinations and divide them into multiple low-correlation subsets based on the minimum support and minimum confidence requirements. For example, purchase records and browsing records are stored in different nodes, and a distributed hash table index is constructed to ensure data query efficiency.Distributed hash table indexes can effectively reduce the risk of data leakage from a single node by distributing data in different nodes, while improving the concurrent performance of data access. For example, in a distributed storage system, each node only stores part of the data, and the target node can be quickly located through the hash table during query, thereby improving system performance while protecting privacy.

[0027] Step S105, continuously optimize the data segmentation and storage mechanism through the collaborative interaction of the federated adversarial learning network and the differential privacy de-correlation processing flow.

[0028] The federated adversarial learning network trains the model locally through multiple participants and uploads the model parameters to the central server for aggregation, avoiding the direct transmission of the original data. In this process, the differential privacy de-correlation processing flow is introduced to reduce the risk of data leakage by adding noise to the model parameters. For example, in the medical data sharing scenario, the electronic medical record data of different hospitals are jointly modeled through the federated learning framework, but when uploading the model parameters, Laplace noise is added to ensure that the sensitive information of specific patients cannot be inferred from the model parameters. This collaborative interaction mechanism not only protects data privacy, but also ensures the training effect of the model. The data segmentation and storage mechanism is another key link in privacy protection. By dividing the data into multiple fragments and storing them on different servers or nodes, even if a node is attacked, the attacker cannot obtain the complete original data. For example, in the financial field, the user's 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 and improves the fault tolerance of the system. When the association risk index exceeds the preset security threshold, the differential privacy de-correlation processing flow triggers an adaptive update. The optimal noise parameter combination is dynamically searched through the immune optimization algorithm to ensure that the impact on model accuracy is minimized while protecting privacy. For example, in the intelligent recommendation system, the user's behavior data is modeled through federated learning. When the system detects that some user behavior data may expose their identity, it will automatically adjust the noise parameters to meet the privacy protection requirements without affecting the accuracy of the recommendation results. When the CNN convolutional neural network model parameters are initialized locally by the participants, the federated learning framework FederatedAveraging ensures the security and consistency of the global model through encrypted aggregation and asynchronous updates. For example, in the field of autonomous driving, the sensor data of different vehicles are trained through the local CNN model, and the model parameters are uploaded to the central server for encrypted aggregation. Due to the use of the asynchronous update mechanism, even if the communication delay of some vehicles is large, 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.

[0029] Step S106, obtaining the original data set and determining a high-correlation field combination, includes: extracting the ID card number and the mobile phone number from the original data set as highly sensitive fields; calculating the Pearson correlation coefficient matrix between the highly sensitive fields and other fields; and determining a high-correlation field combination whose correlation coefficient is greater than a preset threshold based on the Pearson correlation coefficient matrix.

[0030] The reason why the ID card number and mobile phone number are extracted from the original data set as highly sensitive fields is that this information has a high degree of personal privacy and uniqueness, and once leaked, it may lead to serious privacy risks. The ID card number usually contains the individual's date of birth, gender, and region information, while the mobile phone number is an important identifier for personal communication and identity authentication. Therefore, it is reasonable to define these fields as highly 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. The Pearson correlation coefficient matrix between highly sensitive fields and other fields is calculated to quantify the linear relationship between them. The value range of the Pearson correlation coefficient is from negative one to positive one. The closer the value is to positive one or negative one, the stronger the linear relationship between the two fields. For example, suppose that in a certain data set, the correlation coefficient between the ID card number and the date of birth is 0.98, indicating that the two are highly correlated. The correlation coefficient between the mobile phone number and the residential address is 0.12, indicating that the two are almost unrelated. By constructing the correlation coefficient matrix, the degree of correlation between highly sensitive fields and other fields can be intuitively displayed. According to the Pearson correlation coefficient matrix, the highly correlated field combinations with correlation coefficients greater than the preset threshold are determined in order to identify field combinations that pose 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, assuming that in a certain data set, the correlation coefficient between the ID card number and the name is 0.85, which exceeds the preset threshold, then the two fields are considered to be a highly correlated field combination. Such a combination requires special attention because the attacker may infer the ID card number from the name, or vice versa. Through the above steps, the correlation between highly sensitive fields and other fields can be systematically identified, providing data support for subsequent privacy protection measures. Calculating the Pearson correlation coefficient matrix can not only quantify the relationship between the fields, but also help 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.

[0031] Step S107, the intelligent segmentation and distributed storage of highly correlated field combinations includes: using the Apriori algorithm to analyze the highly correlated field combinations; intelligently segmenting the highly correlated field combinations according to minimum support and minimum confidence requirements; distributing and storing the segmented data in different data nodes; and constructing a distributed hash table index based on the distributed stored data.

[0032] To use the Apriori algorithm to analyze highly correlated field combinations, we first need to understand the core idea of ​​the Apriori algorithm. The Apriori algorithm is a classic algorithm for mining frequent item sets and association rules. It finds frequent item sets that meet the minimum support from the data set by searching layer by layer. In this scenario, highly correlated field combinations can be regarded as frequent item sets. The algorithm will gradually generate candidate item sets by scanning the data set multiple times, and filter out combinations whose support exceeds the preset threshold. For example, assuming there is a data set containing fields such as ID card number, mobile phone number, and address, the algorithm will first calculate the support of each field, then generate the support of pairwise combinations, and finally determine which combinations are highly correlated. Intelligently splitting highly correlated field combinations according to the minimum support and minimum confidence requirements is the key to ensuring data dispersion and privacy protection. The minimum support refers to the lowest value that a field combination must reach in the entire data set, while the minimum confidence refers to the ability of the appearance of one field to predict the appearance of another field in a field combination. By setting these two thresholds, field combinations that are truly highly correlated can be filtered out. For example, assuming that the minimum support is set to 0.1 and the minimum confidence is set to 0.8, the algorithm will screen out combinations that appear more than 10% of the time in the data set and have a correlation of more than 80% between fields, and then split these combinations to ensure that sensitive information is not stored in a centralized manner. The split data is stored in different data nodes in a decentralized manner. This step is to achieve distributed storage of data and enhance the scalability of the system. By splitting the highly correlated field combination into multiple parts and storing them on different data nodes, the data load of a single node can be effectively reduced and the processing efficiency of the system can be improved. For example, assuming that the combination of ID card number and mobile phone number is split into two parts, the ID card number is stored in node A, and the mobile phone number is stored in node B. In this way, even if a node is attacked, the attacker cannot obtain complete sensitive information, thereby enhancing data security. Constructing a distributed hash table index based on decentralized stored data is a step to achieve fast query and efficient management of data. A distributed hash table is an index structure used to quickly find data in a distributed system. It maps data to different nodes through a hash function and maintains a global index table. For example, assuming that the segmented data of ID card number and mobile phone number are stored on nodes A and B respectively, the distributed hash table will record the storage location of each field. When a certain field needs to be queried, the system can quickly locate the corresponding node, thereby improving the efficiency of data query. Through the above steps, intelligent segmentation and decentralized storage of highly correlated field combinations can be achieved, which not only protects the privacy of sensitive data, but also improves the processing efficiency and scalability of the system.The core of this process is to screen out highly correlated field combinations through the Apriori algorithm, and to achieve efficient data management through distributed storage and hash table indexes, ultimately achieving the dual goals of data security and system performance.

[0033] Step S108, the encrypted aggregation and asynchronous update of global model parameters are achieved through the federated learning framework, including: initializing the CNN convolutional neural network model parameters locally on the participating party; implementing the encrypted aggregation of the CNN convolutional neural network model parameters through the FederatedAveraging framework; and asynchronously updating the CNN convolutional neural network model parameters.

[0034] When the participants initialize the parameters of the CNN convolutional neural network model locally, each participant independently builds the model structure and initializes the parameters according to its own local data distribution. For example, participants A and B have different image data sets, A's data set is mainly cats and dogs, and B's data set is mainly 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 retain 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, thus laying the foundation for subsequent federated learning. When the encrypted aggregation of CNN convolutional neural network model parameters is implemented through the FederatedAveraging framework, each participant uploads the model parameters to the central server after completing local training. For example, after participants A and B have trained for 10 rounds respectively, they 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 computing technology to encrypt and aggregate the parameters to ensure that the original parameters of a single participant 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 not only protects data privacy, but also optimizes the global model. When asynchronously updating the parameters of the CNN convolutional neural network model, each participant performs local updates based on its own computing resources and data characteristics after receiving the global model parameters. For example, participant A has sufficient computing resources and can perform 5 rounds of local training immediately 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 flexibly adjust the training rhythm according to their own conditions to avoid affecting the overall training efficiency due to the delay of one party. At the same time, asynchronous updates can also effectively cope with network fluctuations and disconnection of participants, ensuring the robustness of the federated learning system. Through the above steps, the federated learning framework can achieve efficient training and optimization of the global model while protecting data privacy. Local initialization ensures the diversity of the model, encrypted aggregation protects data privacy, and asynchronous updates improve the flexibility and robustness of the system. 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-party collaboration.

[0035] Step S109, constructing adversarial samples for high-correlation field combinations and training a privacy protection defense model, includes: constructing adversarial samples for Member reasoning attacks based on the high-correlation field combinations; introducing gradient penalty terms in the training process; and training the privacy protection defense model through the adversarial samples and gradient penalty terms.

[0036] To construct adversarial samples for highly correlated field combinations and train privacy protection defense models, we first need to clarify the definition of highly correlated field combinations and their risks in privacy leakage. Highly correlated field combinations usually refer to fields with strong statistical correlation in the data set, such as ID card number and mobile phone number, name and address, etc. Once these field combinations are exploited by attackers, sensitive information may be restored through reasoning attacks. Therefore, the purpose of constructing adversarial samples for Member reasoning attacks is to simulate the attacker's behavior and generate data samples that can confuse or interfere with the attacker's reasoning ability. For example, when constructing adversarial samples, we can make small perturbations to highly correlated fields, such as randomly replacing certain digits of the mobile phone number, thereby reducing its correlation with the ID card number. This perturbation can not only preserve the overall distribution characteristics of the data, but also effectively prevent attackers from restoring sensitive information through correlation reasoning. The introduction of gradient penalty terms in the training process is to further optimize the robustness of the privacy protection defense model. The core idea of ​​the gradient penalty term is to prevent the model from overfitting the highly correlated features in the training data by constraining the gradient changes of the model during the training process, thereby reducing the vulnerability of the model in reasoning attacks. For example, during the training process, 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 the prediction accuracy but also the smoothness of the gradient during the optimization process. This method can effectively prevent the model from experiencing a significant performance drop under adversarial sample attacks, while improving the model's generalization ability for highly correlated field combinations. The ultimate goal of training a privacy-preserving defense model through adversarial samples and gradient penalty items is to build a model that can effectively resist Member inference attacks in real scenarios. During the training process, the role of adversarial samples is to simulate the attacker's behavior, while the gradient penalty item is used to constrain the optimization direction of the model. For example, during the training process, adversarial samples can be mixed with original data and input into the model, while the gradient penalty item is introduced as an additional loss function. This method enables the model to learn how to maintain stable performance under adversarial sample attacks during the training process, and prevents the model from over-relying on highly correlated field combinations through gradient penalty items. Finally, the trained model can not only perform well on regular data, but also show stronger robustness and privacy protection capabilities when facing Member inference attacks.

[0037] Step S1010, performing differential privacy perturbation processing on the highly sensitive field, includes: using the maximum mean difference method to perform feature transformation on the original data; generating differential privacy random noise based on the Laplace mechanism; and using the differential privacy random noise to perform perturbation processing on the highly sensitive field.

[0038] The maximum mean difference method is a feature transformation technique that aims to extract significant features by calculating the mean difference of data, thereby reducing the sensitivity of data. In the specific implementation, the mean of highly sensitive fields in the original data (such as ID card number and mobile phone number) is first calculated, and then the significant features are determined by comparing the mean difference of these fields with other fields. For example, for the mobile phone number field, the mean of its numbers can be calculated and compared with the mean of other fields such as age and gender to extract key information that can reflect user characteristics. This transformation can effectively reduce the sensitivity of data while retaining the useful information of the data, laying the foundation for subsequent differential privacy processing. Differential privacy random noise generation based on the Laplace mechanism is a classic privacy protection technology. Its core idea is to cover sensitive data by adding random noise that conforms to the Laplace distribution. The characteristic of the Laplace distribution is that it can adjust the intensity of the noise according to the sensitivity of the data, so as to minimize the impact on data accuracy while protecting privacy. For example, for the mobile phone number field, assuming its sensitivity is 10, the intensity of the generated Laplace noise will be adjusted according to this sensitivity to ensure that the added noise can protect privacy without significantly affecting the overall distribution of the data. This mechanism can effectively prevent attackers from inferring users' sensitive information through data analysis. Using differential privacy random noise to perturb highly sensitive fields is the core step of differential privacy protection. In the specific implementation, the generated Laplace noise is added to the highly sensitive field to achieve data perturbation. For example, for the mobile phone number field, assuming that the original value is "12345678910" and the generated noise is "3", the perturbed value is "12345678913". This perturbation processing can effectively prevent attackers from obtaining users' real information through data matching or reasoning, while ensuring that the statistical characteristics of the data are still available. In this way, the normal progress of data analysis tasks can be supported while protecting user privacy. The technical effects of the above method are mainly reflected in the following aspects: First, the maximum mean difference method reduces the sensitivity of the data through feature transformation, which provides a basis for subsequent differential privacy processing; second, the Laplace mechanism can generate appropriate noise according to the sensitivity of the data to ensure the effectiveness of privacy protection; finally, using noise to perturb highly sensitive fields can preserve the availability of data while protecting privacy. These methods together form a strict privacy protection system that can effectively deal with the risks of data leakage and privacy infringement.

[0039] Step S1011, the real-time calculation of the association risk index of the high-correlation field combination and the generation of the association risk analysis report, includes: calculating the association risk index of the high-correlation field combination based on the information entropy principle; generating a data association risk analysis report according to the association risk index; and using the data association risk analysis report as a basis for subsequent dynamic adjustments.

[0040] To calculate the association risk index of a combination of highly correlated fields based on the principle of information entropy, we first need to analyze the fields in the data nodes to identify which fields have high associations. Information entropy is an indicator to measure data uncertainty. By calculating the information entropy of a field, we can evaluate the degree of dependence between fields. For example, in a data set containing the fields "age" and "income", if the information entropy of the "age" field is low, it means that the values ​​of the field are relatively concentrated, while the information entropy of the "income" field is high, indicating that its values ​​are relatively dispersed. By analyzing the joint information entropy of the two fields, we can determine the association between them. If the joint information entropy is significantly lower than the sum of the individual information entropies, it means that there is a high association between the two fields. Furthermore, we can quantify the risks that may be caused by this high association by calculating the association risk index. For example, if the association risk index of the "age" and "income" fields is high, it means that the combination of the two fields may leak sensitive information, such as a high-income group in a specific age group. To generate a data association risk analysis report based on the association risk index, we need to compare the calculated association risk index with the preset security threshold and generate the corresponding analysis results. For example, if the correlation risk index of the "age" and "income" fields is 0.85, and the preset safety threshold is 0.7, then this field combination is considered to have a high correlation risk. These highly correlated field combinations will be listed in detail in the report, and their correlation risk index will be marked. In addition, the report can also include detailed information such as the distribution of fields and correlation strength for subsequent analysis. For example, the report can point out that the values ​​of the "age" field are concentrated between 30 and 40 years old, while the values ​​of the "income" field are between 5,000 and 10,000 yuan. The correlation strength of these two fields is 0.9, indicating that there is a strong dependency between them. The data correlation risk analysis report is used as the basis for subsequent dynamic adjustments. The data segmentation granularity and segmentation dimension can be adjusted according to 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 the "occupation" field can be added as a new segmentation dimension to reduce the correlation between the fields. After the adjustment, the distribution balance index of the field needs to be recalculated to ensure that the segmented data remains balanced in all dimensions. For example, if the data is divided into two subsets according to the "age" field: 30-35 years old and 35-40 years old, it is necessary to check whether the distribution of the "income" field in these two subsets is balanced. If the values ​​of the "income" field in a subset are too concentrated, it is necessary to further adjust the segmentation strategy. In this way, the data segmentation plan can be dynamically optimized to reduce the risks brought by the combination of highly correlated fields.

[0041] Step S1012, dynamically adjusting the data segmentation strategy and optimizing the local model training sample selection according to the associated risk analysis report, includes: adjusting the data segmentation granularity and segmentation dimension according to the associated risk analysis report; recalculating the field distribution balance index; and optimizing the local model training sample selection strategy through a transfer learning method.

[0042] To dynamically adjust the data segmentation strategy and optimize the selection of local model training samples based on the association risk analysis report, first, you need to adjust the granularity and dimension of data segmentation based on the data association evaluation results in the association risk analysis report. For example, in a financial scenario, if it is found that the association risk between user ID and transaction amount is high, the data segmentation granularity can be adjusted from the original segmentation by user ID to segmentation by transaction time, and the segmentation dimension of transaction amount can be increased to reduce data association. In specific implementation, the most appropriate strategy can be selected by analyzing the data distribution characteristics under different segmentation granularities and dimensions. For example, after segmenting the data 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 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 data association. In specific implementation, the distribution of fields under different segmentation strategies can be statistically analyzed to calculate the balance index and select the optimal strategy. For example, by calculating the variance or standard deviation of a field under different segmentation strategies, evaluating its distribution balance, and selecting the strategy with the smallest variance as the final solution. The model performance can be further improved by optimizing the local model training sample selection strategy through transfer learning methods. For example, in a certain image classification task, if there are few local data samples, features can be extracted from pre-trained models in other related fields through transfer learning to optimize the selection of training samples for the local model. In specific implementation, features can be first extracted from the pre-trained model, and then the samples most relevant to the local task can be selected for training based on 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 can not only reduce the risk of data association, but also improve the balance of data segmentation and the efficiency of model training. For example, in a certain 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 has increased by 5%, while the risk of data leakage has decreased by 10%. This technical effect is of great significance in practical applications, and can effectively protect user privacy and improve model performance.

[0043] The above embodiments are only used to illustrate the technical solution of the present invention but not to limit it. The present invention is described in detail with reference to the preferred embodiments. It should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, and should be included in the scope of the claims of the present invention.

Claims

1. A data privacy protection method, characterized in that: include: Obtain the original data set, extract highly sensitive fields and calculate their correlation with other fields, determine the combination of highly correlated fields; intelligently segment and disperse the storage of the highly correlated field combinations, and build a distributed hash table index; realize the encrypted aggregation and asynchronous update of global model parameters through the federated learning framework; construct adversarial samples for the highly correlated field combinations, and train the privacy protection defense model; perform differential privacy perturbation processing on the highly sensitive fields; Calculate the correlation risk index of the highly correlated field combination in real time and generate a correlation risk analysis report; Dynamically adjust the data segmentation strategy according to the associated risk analysis report to optimize the selection of local model training samples; If the association risk index exceeds the preset security threshold, the adaptive update of the differential privacy de-association processing flow is triggered; through the collaborative interaction of the federated adversarial learning network and the differential privacy de-association processing flow, the data segmentation and storage mechanism is continuously optimized.

2. The method according to claim 1, characterized in that The method of obtaining the original data set and determining a high-correlation field combination includes: extracting the ID card number and the mobile phone number from the original data set as highly sensitive fields; calculating the Pearson correlation coefficient matrix between the highly sensitive fields and other fields; and determining a high-correlation field combination whose correlation coefficient is greater than a preset threshold value based on the Pearson correlation coefficient matrix.

3. The method according to claim 1, characterized in that The highly correlated field combinations are intelligently segmented and dispersedly stored, including: using an Apriori algorithm to analyze the highly correlated field combinations; intelligently segmenting the highly correlated field combinations according to minimum support and minimum confidence requirements; distributing and storing the segmented data in different data nodes; and constructing a distributed hash table index based on the dispersedly stored data.

4. The method according to claim 1, characterized in that The encrypted aggregation and asynchronous update of global model parameters are achieved through the federated learning framework, including: initializing the CNN convolutional neural network model parameters locally on the participating party; implementing the encrypted aggregation of the CNN convolutional neural network model parameters through the FederatedAveraging framework; and asynchronously updating the CNN convolutional neural network model parameters.

5. The method according to claim 1, characterized in that Constructing adversarial samples for the highly correlated field combinations and training a privacy protection defense model, including: constructing adversarial samples for Member reasoning attacks based on the highly correlated field combinations; introducing gradient penalty terms during the training process; and training the privacy protection defense model through the adversarial samples and gradient penalty terms.

6. The method according to claim 1, characterized in that The highly sensitive field is subjected to differential privacy perturbation processing, including: using the maximum mean difference method to perform feature transformation on the original data; generating differential privacy random noise based on the Laplace mechanism; and using the differential privacy random noise to perform perturbation processing on the highly sensitive field.

7. The method according to claim 1, characterized in that The real-time calculation of the association risk index of the high-association field combination and the generation of the association risk analysis report include: calculating the association risk index of the high-association field combination based on the information entropy principle; generating a data association risk analysis report according to the association risk index; and using the data association risk analysis report as a basis for subsequent dynamic adjustments.

8. The method according to claim 1, characterized in that Dynamically adjust the data segmentation strategy according to the associated risk analysis report and optimize the local model training sample selection, including: adjusting the data segmentation granularity and segmentation dimension according to the associated risk analysis report; recalculating the field distribution balance index; and optimizing the local model training sample selection strategy through the transfer learning method.

Images