Big data privacy protection analysis system based on federal learning

Abstract

The invention relates to the technical field of big data analysis and privacy protection, and discloses a federated learning-based big data privacy protection analysis system, which comprises a distributed data node cluster, a federated coordination node, an encryption communication module, a privacy enhanced federated learning module and a result verification and output module, the distributed data node cluster comprises a plurality of heterogeneous data nodes, and each data node is provided with a data preprocessing sub-module and a data format conversion sub-module and stores local original big data. According to the method, through a five-layer privacy protection mechanism of data preprocessing de-identification, hybrid encryption communication, parameter homomorphic encryption, differential privacy disturbance and desensitization output, full-process privacy protection from data storage, preprocessing, transmission, model training to result output is realized, and original data leakage and model parameter reverse reasoning risks are effectively avoided; and the requirements of related laws and regulations on data privacy protection are completely met.

Description

Technical Field

[0001] This invention relates to the field of big data analysis and privacy protection technology, specifically a big data privacy protection analysis system based on federated learning. Background Technology

[0002] With the rapid development of the digital economy, big data collaborative analysis has become a core supporting tool in fields such as financial risk control, medical diagnosis, and government decision-making. However, the process of cross-institutional aggregation of raw data poses an extremely high risk of privacy breaches. Sensitive data such as personal identity information, health records, and transaction records are vulnerable to theft, alteration, or misuse once centrally stored or transmitted, which not only infringes on users' legitimate rights and interests but may also violate the mandatory requirements of laws and regulations such as the Personal Information Protection Law, the Data Security Law, and the Cybersecurity Law.

[0003] In existing technologies, big data privacy protection solutions mainly fall into two categories: one is traditional privacy protection methods, including encrypted data storage and static anonymization. Encrypted storage only guarantees security in the static state of the data; when the data is decrypted for analysis, it still faces the risk of privacy leakage. Traditional anonymization (such as deleting ID numbers, phone numbers, and other identifying information) is easily exploited by attackers to reconstruct the original data through correlation analysis (combining publicly available data or other dimensional information), thus offering limited privacy protection. The other category is collaborative analysis solutions based on federated learning. Its core idea is "the model moves while the data remains stationary," achieving data value mining through collaborative training of a global model by various participants. However, existing federated learning systems still have many technical shortcomings. Insufficient communication security: Model parameter transmission often uses a single encryption mechanism or does not perform high-strength encryption during the transmission process, making it easy to be intercepted and cracked, resulting in the leakage of original data features through reverse inference from model parameters; Inadequate privacy enhancement mechanisms: The model aggregation process lacks dynamic adaptation privacy protection strategies, fixed privacy parameters cannot meet the protection needs of data with different sensitivity levels, and the aggregation parameters are not effectively perturbed, posing a risk of reverse inference. Poor adaptability to heterogeneous data: It is difficult to coordinate the analysis of structured data (such as database tables), semi-structured data (such as XML and JSON files) and unstructured data (such as text and images). Differences in data formats lead to low model training efficiency and poor accuracy. Lack of anomaly monitoring and fault tolerance: There is a lack of real-time monitoring of the running status of data nodes, data transmission process and model training behavior, making it impossible to identify abnormal situations such as malicious node attacks and hardware failures in a timely manner. Furthermore, there is no effective data completion and model training compensation mechanism after an abnormal node is removed, which affects system stability. Rigid model optimization mechanism: The hyperparameters of the global model are mostly fixed values ​​and are not dynamically adjusted according to the data distribution characteristics of each node, resulting in slow model convergence speed and insufficient analysis accuracy.

[0004] Therefore, developing a federated learning system with multi-layered privacy protection mechanisms, adaptability to heterogeneous data, support for dynamic model optimization, and both anomaly monitoring and fault tolerance capabilities has become the key to solving the pain points of existing technologies. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a big data privacy protection analysis system based on federated learning, which solves the problems mentioned in the background section.

[0006] (II) Technical Solution To achieve the above objectives, this invention provides the following technical solution: a big data privacy protection and analysis system based on federated learning, comprising a distributed data node cluster, a federated coordination node, an encrypted communication module, a privacy-enhanced federated learning module, a result verification and output module, and an anomaly monitoring module. These modules work collaboratively to complete the privacy protection and collaborative analysis of big data. The specific technical solution is as follows: 1. Distributed Data Node Cluster: Contains at least two heterogeneous data nodes, each corresponding to an independent data source (such as financial institutions, medical institutions, government departments, etc.), storing local raw big data. Each data node integrates a data preprocessing submodule and a data format conversion submodule. The data preprocessing submodule employs a three-stage processing flow: de-identification, outlier removal, and data standardization. De-identification includes deleting sensitive personal identifiers (such as names, ID numbers, and hospital numbers) and desensitizing identifiable identifiers (such as phone numbers, addresses, and email addresses) using SHA-256 hash encryption. Outlier removal uses a box plot method, calculating the data quartiles (Q1, Q3) to determine the outlier threshold (Q3 + 1.5IQR to Q1 - 1.5IQR, IQR = Q3 - Q1), identifying and removing outliers exceeding the threshold range. Data standardization uses a Min-Max normalization algorithm to map the data to the [0,1] interval, with the formula: Where x is the original data, x min x is the minimum value of the data. max This represents the maximum value of the data.

[0007] The data format conversion submodule employs a unified tensor conversion protocol to convert heterogeneous data into tensor formats with consistent dimensions. For structured data (such as database tables), it is directly converted to two-dimensional tensors; for semi-structured data (such as XML and JSON files), core features are extracted by parsing key-value pairs and then converted to two-dimensional tensors; for unstructured data (such as text and images), text data is converted to word vector tensors (dimension [number of samples, 768]) using the BERT model, and image data is converted to feature tensors (dimension [number of samples, 2048]) after feature extraction using a CNN model, ensuring that different types of data can participate in unified model training.

[0008] Each data node only outputs homomorphically encrypted local model parameters to the federation coordination node, without leaking any original data or unencrypted model parameter information.

[0009] 2. Federation Coordination Node: As the core control node of the system, it integrates the model initialization submodule, model aggregation submodule, and model optimization submodule. Model initialization submodule: Select the appropriate base model according to the analysis task type, including logistic regression model (suitable for classification tasks), random forest model (suitable for feature importance analysis), neural network model (suitable for complex pattern recognition), and initialize the initial values ​​of model hyperparameters (learning rate initial value is 0.01, number of iterations initial value is 100, number of hidden layer neurons initial value is 64).

[0010] The model aggregation submodule receives encrypted model parameters uploaded from each data node and aggregates them using a weighted average algorithm. The weights are determined based on the proportion of samples from each data node, as shown in the formula: W global N represents the global model parameters. i Let N be the number of samples in the i-th data node. total W represents the total number of samples across all data nodes. i These are the encrypted local model parameters for the i-th data node.

[0011] The model optimization submodule dynamically adjusts the global model hyperparameters based on the Bayesian optimization algorithm, aiming to minimize the model's loss function value. The Bayesian optimization algorithm uses a Gaussian process as the prior distribution, with 50 iterations. In each iteration, the posterior distribution is updated based on the model performance (loss function value, accuracy) of the preceding hyperparameter combinations, outputting the optimal hyperparameter combination. The adjusted hyperparameters include the learning rate (range 0.001-0.1), the number of iterations (range 50-200), and the number of hidden layer neurons (range 32-128), ensuring the model adapts to the data distribution characteristics of each node.

[0012] 3. Encrypted Communication Module: Employs a hybrid encryption mechanism combining asymmetric and symmetric encryption to achieve encrypted bidirectional data transmission between the distributed data node cluster and the federated coordination node. Key negotiation process: The RSA-2048 asymmetric encryption algorithm is used. The federation coordination node generates a public key PK and a private key SK, and distributes the public key PK to each data node. Each data node generates a random AES-256 symmetric key K, encrypts K with the public key PK, and sends it to the federation coordination node. The federation coordination node decrypts K with the private key SK to obtain K, thus completing the key negotiation.

[0013] Data transmission encryption: The transmitted data (including model parameters, node state information, etc.) is encrypted using the CBC mode of the AES-256 algorithm. The initialization vector (IV) is randomly generated and transmitted along with the encrypted data to ensure that the encryption result is different for each transmission.

[0014] The key dynamic update submodule periodically updates the symmetric encryption key K based on timestamps and unique node identifiers (NodeIDs). The update cycle is set to 2 hours (experiments have verified that this cycle balances security and transmission efficiency, avoiding system overhead caused by frequent updates). The update process is as follows: the federated coordination node generates a new key K′, encrypts K′ with the original key K, and distributes it to each data node. Each node decrypts the key and updates its local key, completing the key rotation.

[0015] 4. Privacy-Enhanced Federated Learning Module: Deployed on each data node and federated coordination node, integrating local training submodules, parameter encryption submodules, global aggregation submodules, and pattern adaptation submodules. Local training submodule: Gradient descent algorithm is used (stochastic gradient descent SGD for classification tasks and adaptive moment estimation ADAM for regression tasks). The local model is trained based on the local preprocessed tensor data, and local model parameters (weights, biases, etc.) are generated. An early stopping strategy is adopted during training (training is stopped when the loss function value on the validation set does not decrease for 5 consecutive iterations) to avoid overfitting.

[0016] The parameter encryption submodule uses the Paillier homomorphic encryption algorithm to encrypt local model parameters. This algorithm supports additive homomorphism and scalar multiplication homomorphism, and can complete model parameter aggregation operations in the encrypted state without decryption, thus ensuring the privacy and security of parameter transmission and aggregation processes.

[0017] The global aggregation submodule introduces a differential privacy mechanism based on weighted average aggregation. This involves adding Laplace noise to perturb the aggregation parameters. The noise intensity is negatively correlated with the privacy budget ε, as shown in the formula: , where Δf is the sensitivity of the aggregation function (Δf is 1.0 in this system), and Lap(0,Δf / ε) is noise that follows a Laplace distribution.

[0018] Pattern adaptation submodule: Automatically selects federated learning mode by calculating the feature similarity and sample overlap between data nodes: when feature similarity ≥ 80% and sample overlap < 20%, select horizontal federated learning mode; when sample overlap ≥ 80% and feature similarity < 20%, select vertical federated learning mode; when feature similarity < 50% and sample overlap < 50%, select federated transfer learning mode (using pre-trained models to initialize local models to improve training efficiency).

[0019] 5. Result Verification and Output Module: Integrates a consistency verification submodule and a de-identification output submodule. The consistency verification submodule verifies the effectiveness of the global model by calculating the mean squared error (MSE) between the analysis results of each local model and the analysis results of the global model. The MSE calculation formula is as follows: , where y i The results are from a local model analysis. This represents the global model analysis results, where n is the number of analyzed samples. The preset MSE threshold is 0.05 (experiments have verified that when MSE ≤ 0.05, the model prediction accuracy is ≥ 95%, meeting the requirements of practical applications). When the MSE exceeds the threshold, global model retraining is triggered; when the MSE is below the threshold for three consecutive iterations, the model training is considered complete.

[0020] The desensitized output submodule employs a three-tiered desensitization strategy: data generalization, anonymization, and sensitive information filtering. Data generalization converts specific values ​​into ranges (e.g., age 28 → 25-30 years old, transaction amount 5000 yuan → 4000-6000 yuan); anonymization removes all personal identification information (including direct and indirect identifiers); sensitive information filtering uses a keyword matching algorithm (the sensitive word library contains over 100 categories of sensitive information keywords such as ID card numbers, mobile phone numbers, and bank card numbers) to automatically remove sensitive content from the analysis report, ultimately outputting a desensitized structured analysis report.

[0021] 6. Anomaly Monitoring Module: As the core of system security protection, it integrates a status acquisition submodule, an anomaly identification submodule, and a fault tolerance processing submodule. Status acquisition submodule: Collects the running status (CPU utilization, memory usage, disk I / O rate), data transmission traffic (bytes per second, number of data packets), and model training process data (training iteration speed, parameter update magnitude, loss function change rate) of each data node in real time with a 1-minute acquisition cycle.

[0022] Anomaly Identification Submodule: Set multi-dimensional anomaly thresholds: CPU utilization exceeding 90% for 10 consecutive minutes, memory usage exceeding 85% for 10 consecutive minutes, data transmission traffic increasing by more than 50% or decreasing by more than 80% compared to the average traffic in the most recent hour, parameter update amplitude exceeding ±50%, loss function value increasing by more than 100%. Meeting any one of these conditions will be judged as abnormal behavior.

[0023] Fault Tolerance Submodule: When abnormal behavior is detected, an alarm message is immediately generated (including abnormal node identifier, abnormal type, abnormal time, and abnormal data indicators) and synchronized to the system administrator via SMS and email. At the same time, the communication connection of the abnormal node is automatically cut off. If the abnormal node is a non-core node (sample percentage <10%), the federated coordination node re-aggregates and updates the global model based on the model parameters of the remaining nodes. If the abnormal node is a core node (sample percentage ≥10%), local model retraining is triggered. After the abnormal node is repaired, it is reconnected to the system through the data synchronization mechanism to ensure continuous system operation.

[0024] (III) Beneficial Effects This invention provides a big data privacy protection analysis system based on federated learning, which has the following beneficial effects: Privacy protection is comprehensive and robust throughout the entire process: Through a five-layer privacy protection mechanism of "data preprocessing de-identification + hybrid encrypted communication + parameter homomorphic encryption + differential privacy perturbation + desensitized output", privacy protection is achieved throughout the entire process from data storage, preprocessing, transmission, model training to result output. This effectively avoids the risk of raw data leakage and reverse inference of model parameters, and fully complies with the relevant laws and regulations on data privacy protection. Strong adaptability to heterogeneous data: Through the unified tensor conversion protocol of the data format conversion submodule, it is compatible with the collaborative analysis of structured, semi-structured and unstructured data. Combined with the automatic federated learning mode switching of the pattern adaptation submodule, it can meet the cross-domain data collaborative analysis needs in different scenarios such as finance-healthcare and government collaboration, and has a wide range of applications. Excellent model analysis accuracy and convergence speed: The model optimization submodule dynamically adjusts hyperparameters based on the Bayesian optimization algorithm, combined with local training early stopping strategy and weighted average aggregation strategy, which improves the global model analysis accuracy by 10%-15% and convergence speed by 20%-30%. At the same time, the result verification mechanism ensures that the model output is valid. The system boasts high security and stability: the real-time monitoring and fault-tolerant processing mechanism of the anomaly monitoring module can quickly identify abnormal behaviors such as malicious attacks and hardware failures. Through fault-tolerant measures such as cutting off abnormal connections, re-aggregating or retraining models, it ensures uninterrupted system operation with a fault recovery time of ≤30 minutes. Balancing transmission efficiency and security: The hybrid encryption mechanism balances the high-efficiency transmission of AES-256 with the secure key negotiation of RSA-2048. The dynamic key update mechanism further reduces the risk of key leakage, and the data transmission latency is ≤50ms, meeting the real-time requirements of large-scale collaborative data analysis. Attached Figure Description

[0025] Figure 1 This is a flowchart of the big data privacy protection analysis system based on federated learning according to the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] Example 1: Cross-domain big data collaborative analysis scenario in finance and healthcare This embodiment is applied to cross-domain collaborative analysis between banking institutions and top-tier hospitals. The goal is to uncover the correlation between users' financial behavior and health status, providing data support for banks' accurate credit assessment and hospitals' health risk early warning, while ensuring the privacy and security of users' financial and health data.

[0028] System Deployment: Distributed data node cluster: including financial data nodes (deployed in bank data centers, storing structured financial data of 1 million users, including deposit balances, loan records, transaction flow, etc., with a sample size of 600,000) and medical data nodes (deployed in hospital data centers, storing semi-structured / unstructured medical data of 800,000 users, including medical record texts, examination indicators, image features, etc., with a sample size of 400,000); Federal coordination nodes: deployed in third-party industry association data centers to ensure the impartiality of collaborative analysis; The encrypted communication module, privacy-enhanced federated learning module, result verification and output module, and anomaly monitoring module are integrated into the above nodes and communicate through a dedicated network.

[0029] Data preprocessing and format conversion: Financial data node: The data preprocessing submodule deletes user names and ID numbers, encrypts and desensitizes mobile phone numbers using SHA-256, uses box plots to remove abnormal data with transaction amounts > 1 million yuan, and normalizes data such as deposit balances and loan amounts to the [0,1] range; the data format conversion submodule directly converts structured financial data into a two-dimensional tensor of [600000,10] dimensions (10 financial features).

[0030] Medical data node: The data preprocessing submodule deletes patient names and hospital numbers, uses box plots to remove abnormal data of examination indicators that exceed the normal range by 3 times, and normalizes indicators such as blood pressure and blood sugar to the [0,1] interval; The data format conversion submodule uses the BERT model to convert medical record text into word vector tensors of [400000,768] dimensions, converts examination indicators into two-dimensional tensors of [400000,8] dimensions, and finally merges them into a unified tensor of [400000,776] dimensions.

[0031] Encrypted communication establishment and key update: Key negotiation: The federal coordination node generates an RSA-2048 public key and a private key, and distributes the public key to the financial data node and the medical data node; the two nodes generate an AES-256 symmetric key, encrypt it with the public key and send it to the federal coordination node, and the federal coordination node decrypts it with the private key to obtain the symmetric key, thus completing the establishment of encrypted communication.

[0032] Key Update: The key dynamic update submodule updates the symmetric key every 2 hours. After the federated coordination node generates a new key, it encrypts and distributes it using the original key. Both nodes decrypt the key and update their local keys to ensure communication security.

[0033] Federated learning model training: Mode selection: The data nodes have a sample overlap of 75% (financial and medical data of the same user) and a feature similarity of 15%. The mode adaptation submodule automatically selects the vertical federated learning mode.

[0034] Model initialization: The federated coordination node selects a logistic regression model as the base model, initializes the learning rate to 0.01, and sets the number of iterations to 100.

[0035] Local training: The local training submodule for financial data nodes and medical data nodes uses the SGD algorithm to train local models and generate local model parameters. During the training process, the early stopping strategy is triggered when the validation set loss does not decrease for 5 consecutive times.

[0036] Parameter encryption and upload: The parameter encryption submodule uses the Paillier homomorphic encryption algorithm to encrypt the local model parameters and then uploads them to the federated coordination node through the encrypted communication module.

[0037] Global aggregation and optimization: The global aggregation submodule aggregates parameters by weighted average based on sample proportion (60% for financial nodes and 40% for medical nodes), and adds Laplace noise (both nodes have a high sensitivity level, and the privacy budget ε=0.3); The model optimization submodule is based on the Bayesian optimization algorithm. After 50 iterations, the learning rate is adjusted to 0.008, the number of iterations is adjusted to 120, the global model parameters are updated, and distributed to the two nodes; the above process is repeated until the model converges (the loss function value is lower than 0.01).

[0038] Result verification and output: Consistency check: The consistency check submodule calculates the MSE of the analysis results of the local model and the global model of the two nodes. If the MSE is 0.03 < 0.05, the global model is determined to be valid.

[0039] Desensitized Output: The desensitized output submodule generalizes credit risk levels into three levels: "low, medium, and high," and health risk assessment results into three levels: "low risk, medium risk, and high risk." It removes all personal identification information, generates cross-domain collaborative analysis reports, and sends them to the bank and hospital respectively.

[0040] Anomaly monitoring and fault tolerance: Anomaly Detection: The anomaly monitoring module collects node status in real time. If the CPU utilization of a medical data node reaches 95% for 10 consecutive minutes (exceeding the threshold of 90%), it is judged as abnormal behavior.

[0041] Anomaly Handling: An alarm message was generated and sent to the system administrator, and the communication connection between the medical data nodes was simultaneously cut off; because the medical data nodes accounted for 40% of the samples (core nodes), local model retraining was triggered; the administrator investigated and found that it was a hardware failure. After repair, the medical data nodes reconnected to the system through the data synchronization mechanism, and the system returned to normal operation.

[0042] In this embodiment, the system does not aggregate any raw financial or medical data. It achieves cross-domain data value mining through collaborative training, and its privacy protection complies with legal and regulatory requirements. The analysis report provides reliable support for bank credit assessment and hospital health early warning, thus verifying the effectiveness and practicality of the system.

[0043] Example 2: Collaborative Analysis Scenario for Government Big Data This embodiment is applied to the collaborative analysis of government big data by the market supervision bureau, tax bureau, and human resources and social security bureau. The goal is to optimize the business environment, improve the efficiency of government services, and protect the privacy and security of government data for enterprises and individuals.

[0044] System deployment: The distributed data node cluster includes market supervision data nodes, tax data nodes, and human resources and social security data nodes, which store enterprise registration information, tax payment data, and social security payment data, respectively; the federal coordination node is deployed on the government cloud platform, and other modules are integrated into each node.

[0045] Key differences: Mode selection: The data feature similarity of each node is 85% (all are business-related features), the sample overlap is 20%, and the mode adaptation submodule automatically selects the horizontal federated learning mode.

[0046] Privacy Budget: The sensitivity level of government data is medium, and the privacy budget ε=1.5.

[0047] Base Model: The model optimization submodule selects the random forest model as the base model, and initializes the hyperparameters as 100 decision trees and 10 maximum tree depth.

[0048] Implementation results: The system efficiently completed collaborative analysis of government big data, with a global model analysis accuracy rate of 96% and MSE=0.02. The anonymized analysis report provides data support for government departments to optimize approval processes and conduct precise supervision, while ensuring the privacy and security of government data.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A big data privacy-preserving analysis system based on federated learning, characterized in that: include: The system includes a distributed data node cluster, a federated coordination node, an encrypted communication module, a privacy-enhanced federated learning module, and a result verification and output module. The distributed data node cluster contains multiple heterogeneous data nodes. Each data node is equipped with a data preprocessing submodule and a data format conversion submodule. It stores local raw big data and only outputs encrypted local model parameters to the outside, without leaking the original data and unencrypted model parameter information. The data preprocessing submodule is used to perform de-identification, outlier removal, and data standardization on the local raw big data. The data format conversion submodule is used to convert heterogeneous data into a unified tensor format, wherein the heterogeneous data includes structured data, semi-structured data, and unstructured data. The federated coordination node is used to initialize global model parameters, receive encrypted model parameters uploaded by each data node, perform global model aggregation update, and distribute the updated global model parameters to each data node. The encrypted communication module adopts a hybrid encryption mechanism to realize bidirectional data transmission encryption between the distributed data node cluster and the federated coordination node. The hybrid encryption mechanism combines an asymmetric encryption algorithm for key negotiation and a symmetric encryption algorithm for data transmission. The privacy-enhanced federated learning module is deployed on each data node and federated coordination node. It is used to complete local model training based on local data and introduce a differential privacy mechanism to perturb the aggregation parameters during the global model aggregation process, thereby achieving privacy protection during the model training process. The result verification and output module is used to verify the validity of the analysis results of the global model and output the final analysis report in an anonymized form.

2. The big data privacy protection analysis system based on federated learning according to claim 1, characterized in that: The de-identification process of the data preprocessing submodule includes deleting sensitive personal identifiers and encrypting and desensitizing aligner identifiers; outlier removal uses box plot method; and data standardization process normalizes the data to the [0,1] interval.

3. The big data privacy protection analysis system based on federated learning according to claim 1, characterized in that: The asymmetric encryption algorithm in the encrypted communication module is the RSA-2048 algorithm, and the symmetric encryption algorithm is the AES-256 algorithm. The encrypted communication module also has a key dynamic update submodule, which periodically updates the symmetric encryption key based on timestamps and node identity information.

4. The big data privacy protection analysis system based on federated learning according to claim 1, characterized in that: The privacy-enhanced federated learning module includes a local training submodule, a parameter encryption submodule, a global aggregation submodule, and a mode adaptation submodule. The local training submodule uses a gradient descent algorithm to train the local model and generate local model parameters. The parameter encryption submodule performs homomorphic encryption on the local model parameters and uploads them to the federated coordination node. The global aggregation submodule uses a weighted average algorithm to aggregate the encrypted local model parameters and adds Laplace noise through a differential privacy mechanism to perturb the parameters. The mode adaptation submodule is used to automatically select horizontal federated learning, vertical federated learning, or federated transfer learning modes based on the data distribution characteristics between data nodes.

5. The big data privacy protection analysis system based on federated learning according to claim 4, characterized in that: The privacy budget ε of the differential privacy mechanism can be dynamically adjusted according to the sensitivity level of the data node. The higher the sensitivity level, the smaller the privacy budget ε and the greater the noise disturbance intensity. Specifically, when the sensitivity level is high, ε∈[0.1,0.5] and when the sensitivity level is low, ε∈[1.0,2.0].

6. The big data privacy protection analysis system based on federated learning according to claim 1, characterized in that: The result verification and output module includes a consistency verification submodule and a desensitization output submodule. The consistency verification submodule verifies the effectiveness of the global model by calculating the mean square error (MSE) between the analysis results of each local model and the analysis results of the global model. When the MSE exceeds a preset threshold, the global model is retrained. The desensitization output submodule uses data generalization and anonymization processing to delete sensitive information in the analysis report.

7. A big data privacy protection analysis system based on federated learning according to claim 6, characterized in that: The preset threshold is 0.

05.

8. The big data privacy protection analysis system based on federated learning according to claim 1, characterized in that, It also includes an anomaly monitoring module, which is used to monitor the operating status, data transmission traffic and model training process of each data node in real time. When an abnormal behavior is detected, an alarm message is generated and the communication connection of the abnormal node is cut off.

9. A big data privacy protection analysis system based on federated learning according to claim 8, characterized in that, The abnormal behaviors include CPU utilization exceeding 90% for 10 consecutive minutes, a sudden increase in data transmission traffic of more than 50%, or a parameter update range exceeding ±50%.

10. A big data privacy protection analysis system based on federated learning according to claim 1, characterized in that, The federated coordination node also has a model optimization submodule, which dynamically adjusts the hyperparameters of the global model based on the Bayesian optimization algorithm and the local data distribution characteristics of each data node.

Images