Sensitive data auditing method based on neural network multilayer perception

Abstract

The invention relates to the technical field of neural network multilayer perception, in particular to a sensitive data auditing method based on neural network multilayer perception, which comprises the steps of data collection, data preprocessing, feature selection, multilayer perception neural network integration, model training, real-time data stream processing, model evaluation, result analysis and privacy protection enhancement. According to the invention, through the deep learning ability of the multi-layer perceptron (MLP) neural network, the perception and extraction ability of the data features is significantly improved. Due to the introduction of a real-time data stream processing technology, the system can respond to sensitive data leakage in time, and the security risk is effectively reduced. The self-adaptive learning mechanism ensures that the model can continuously adapt to novel threats, and the efficient detection capability is kept. The differential privacy technology provides powerful privacy protection in the data processing and model training process, and the risk of user privacy disclosure is avoided.

Description

Technical Field

[0001] The present invention relates to the field of neural network multi-layer perception technology, and in particular to a sensitive data auditing method based on neural network multi-layer perception. Background Art

[0002] In today's information age, data has become the core asset of enterprises and organizations. However, with the rapid growth of data volume and the continuous development of data processing technology, sensitive data leakage incidents have also occurred frequently, bringing huge risks and losses to enterprises and users. Traditional sensitive data audit methods mainly rely on static rules and feature engineering, and detect abnormal behaviors and data leakage through manually set rules. These methods are powerless in the face of increasingly complex attack patterns and massive amounts of data, and it is difficult to provide timely and accurate audit results. When processing large amounts of data, traditional sensitive data audit methods usually rely on offline analysis and cannot detect and respond to data leakage incidents in real time, which may cause significant losses before security threats are detected and processed. In addition, traditional methods rely on static rules and feature engineering, which are difficult to deal with new and unknown threats, which makes them show low detection rate and accuracy when facing ever-changing attack patterns. Due to the lack of effective privacy protection mechanisms, traditional methods are prone to user privacy leakage during data processing, further increasing the risk of data security. These deficiencies limit the effectiveness and adaptability of traditional sensitive data audit methods, making it difficult to provide sufficient security protection in modern complex network environments. Summary of the invention

[0003] The purpose of the present invention is to solve the technical problems raised in the above background technology.

[0004] The present invention adopts the following technical solution: a sensitive data audit method based on neural network multi-layer perception, comprising the following steps:

[0005] S1: Data collection, collecting all relevant data in the target system, including log files, user behavior records, database content, etc.;

[0006] S2: Data preprocessing: cleaning and preprocessing the collected data, removing noise data, filling missing values, and standardizing the data;

[0007] S3: Feature selection, extracting meaningful features from preprocessed data through feature engineering for subsequent model training and analysis;

[0008] S4: Multi-layer perceptron neural network integration, using multi-layer perceptron (MLP) neural network to conduct deep learning on data, and improve the perception ability and accuracy of the model through multi-layer nonlinear transformation;

[0009] S5: Model training, using traditional machine learning algorithms (such as logistic regression, decision trees, etc.) to model the data and train the model to identify potential sensitive data leakage or abnormal behavior;

[0010] S6: Real-time data stream processing, integrating with real-time data stream processing frameworks (such as Apache Kafka or Flink), so that sensitive data audits can be performed in real time instead of relying on offline data analysis;

[0011] S7: Model evaluation: evaluate the trained model through cross-validation and other evaluation methods to ensure that its performance and accuracy meet the requirements;

[0012] S8: Result analysis: Analyze the results of the model output, identify and report possible sensitive data leakage incidents, and propose corresponding security measures;

[0013] S9: Enhanced privacy protection. Differential privacy technology is used in data processing and model training to ensure that user privacy is not leaked during sensitive data audits, thereby improving the security and trust of the system.

[0014] Preferably, the step S1 includes the following steps:

[0015] S101: Log file acquisition, extracting all relevant log files from the target system, including system logs, security logs, application logs, etc., to ensure full coverage of system activities;

[0016] S102: User behavior record collection, by monitoring and recording user operation behaviors, collecting user access records, operation logs, login records, etc., for subsequent behavior analysis;

[0017] S103: Database content extraction, obtaining all necessary data tables and fields from the database, including structured and unstructured data, to ensure data comprehensiveness and integrity.

[0018] Preferably, the step S2 includes the following steps:

[0019] S201: Data cleaning, ensuring data accuracy and consistency by removing duplicate data, correcting erroneous data, and deleting irrelevant data;

[0020] S202: Missing value processing: filling or deleting missing data using commonly used methods such as mean filling, interpolation, or model-based prediction filling;

[0021] S203: Data standardization: standardize the numerical data to make it conform to a unified scale. Common methods include z-score standardization and min-max normalization.

[0022] Preferably, S3 includes the following steps:

[0023] S301: Feature extraction: extracting information that can represent data features from raw data through feature engineering. Methods used include principal component analysis (PCA), linear discriminant analysis (LDA), etc.

[0024] S302: Feature screening, using correlation analysis, information gain and other methods to screen out the most representative features and remove redundant and noisy features;

[0025] S303: Feature conversion: reconstruct the features through feature conversion techniques such as polynomial conversion and frequency domain transformation to improve the expressiveness of the features and the generalization ability of the model.

[0026] Preferably, the step S4 comprises the following steps:

[0027] S401: Model architecture design, design the architecture of the multi-layer perceptron (MLP) neural network, including the number of nodes and connection methods of the input layer, multiple hidden layers, and output layer;

[0028] S402: Initialize network parameters, initialize the weights and biases of the neural network, and use methods such as Xavier initialization and He initialization to ensure the stability and convergence speed of network training;

[0029] S403: Activation function selection: select an appropriate activation function for each layer, such as ReLU, Sigmoid, Tanh, etc., to improve the nonlinear expression ability and training efficiency of the model.

[0030] Preferably, the step S5 includes the following steps:

[0031] S501: training set and test set division: divide the preprocessed data into training set and test set, usually using cross-validation or holdout method to ensure fairness in model training and evaluation;

[0032] S502: Hyperparameter adjustment: Tune the model's hyperparameters, such as learning rate, batch size, regularization parameters, etc., through grid search, random search, or Bayesian optimization.

[0033] S503: Gradient descent optimization, using gradient descent algorithms (such as SGD, Adam, etc.) to optimize the parameters of the model, and through multiple iterative training, minimize the loss function and improve the accuracy of the model.

[0034] Preferably, the step S6 comprises the following steps:

[0035] S601: Build a data stream architecture and a real-time data stream processing framework (such as Apache Kafka or Flink) to achieve real-time data collection, transmission, and processing;

[0036] S602: Data flow monitoring: real-time monitoring of the status and performance of data flows to ensure the stability and efficiency of data flow processing and timely detection and resolution of abnormal situations;

[0037] S603: Real-time analysis and processing: Integrate real-time analysis algorithms to process and analyze real-time data streams, and promptly detect and respond to potential sensitive data leakage incidents.

[0038] Preferably, the step S7 includes the following steps:

[0039] S701: Evaluation indicator selection: select appropriate evaluation indicators, such as accuracy, recall, F1-score, ROC-AUC, etc., to comprehensively evaluate the model performance;

[0040] S702: Cross-validation: Use the cross-validation method to evaluate the stability and generalization ability of the model through multiple training and validation to prevent overfitting;

[0041] S703: Model tuning: Based on the evaluation results, the model is tuned and improved by adjusting hyperparameters, increasing or decreasing the number of network layers, improving feature engineering, etc., to improve model performance.

[0042] Preferably, the step S8 includes the following steps:

[0043] S801: Anomaly detection, by analyzing the output of the model, identifying potential abnormal behaviors and sensitive data leakage events, and performing detailed classification and labeling;

[0044] S802: Security incident report: Generate a security incident report to record in detail the sensitive data leakage incidents found, including time, location, incident description and scope of impact;

[0045] S803: Security measures recommendations: Based on the analysis results, corresponding security measures and improvement suggestions are proposed, such as strengthening authority management, increasing data encryption strength, improving log monitoring, etc.

[0046] Preferably, the step S9 includes the following steps:

[0047] S901: Differential privacy mechanism design: Introduce differential privacy mechanism in data processing and model training, protect user privacy and prevent sensitive information leakage by adding noise and other methods;

[0048] S902: Privacy protection evaluation: evaluate the effectiveness of the differential privacy mechanism to ensure that it can provide adequate privacy protection while ensuring data availability;

[0049] S903: Privacy policy formulation: formulate and implement privacy protection policies to ensure strict compliance with privacy protection laws, regulations and industry standards during data collection, processing, analysis and storage.

[0050] Compared with the prior art, the advantages and positive effects of the present invention are:

[0051] In the present invention, the sensitive data auditing method based on neural network multi-layer perception significantly improves the perception and extraction capabilities of data features through the deep learning capabilities of the multi-layer perceptron (MLP) neural network. The introduction of real-time data stream processing technology enables the system to respond to sensitive data leaks in a timely manner, effectively reducing security risks. The adaptive learning mechanism ensures that the model can continuously adapt to new threats and maintain efficient detection capabilities. Differential privacy technology provides strong privacy protection during data processing and model training, avoiding the risk of user privacy leakage. The combination of these technical means makes sensitive data auditing more accurate, efficient and secure, and can effectively cope with complex and changing network security environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of a sensitive data auditing method based on neural network multi-layer perception proposed in the present invention. DETAILED DESCRIPTION

[0053] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0054] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments of the following disclosure.

[0055] Implementation Methods

[0056] See also Figure 1 The present invention provides a technical solution: a sensitive data audit method based on neural network multi-layer perception, comprising the following steps:

[0057] S1: Data collection, collecting all relevant data in the target system, including log files, user behavior records, database content, etc.;

[0058] S2: Data preprocessing: cleaning and preprocessing the collected data, removing noise data, filling missing values, and standardizing the data;

[0059] S3: Feature selection, extracting meaningful features from preprocessed data through feature engineering for subsequent model training and analysis;

[0060] S4: Multi-layer perceptron neural network integration, using multi-layer perceptron (MLP) neural network to conduct deep learning on data, and improve the perception ability and accuracy of the model through multi-layer nonlinear transformation;

[0061] S5: Model training, using traditional machine learning algorithms (such as logistic regression, decision trees, etc.) to model the data and train the model to identify potential sensitive data leakage or abnormal behavior;

[0062] S6: Real-time data stream processing, integrating with real-time data stream processing frameworks (such as Apache Kafka or Flink), so that sensitive data audits can be performed in real time instead of relying on offline data analysis;

[0063] S7: Model evaluation: evaluate the trained model through cross-validation and other evaluation methods to ensure that its performance and accuracy meet the requirements;

[0064] S8: Result analysis: Analyze the results of the model output, identify and report possible sensitive data leakage incidents, and propose corresponding security measures;

[0065] S9: Enhanced privacy protection. Differential privacy technology is used in data processing and model training to ensure that user privacy is not leaked during sensitive data audits, thereby improving the security and trust of the system.

[0066] Furthermore, S1 includes the following steps:

[0067] S101: Log file acquisition, extracting all relevant log files from the target system, including system logs, security logs, application logs, etc., to ensure full coverage of system activities;

[0068] S102: User behavior record collection, by monitoring and recording user operation behaviors, collecting user access records, operation logs, login records, etc., for subsequent behavior analysis;

[0069] S103: Database content extraction, obtaining all necessary data tables and fields from the database, including structured and unstructured data, to ensure data comprehensiveness and integrity.

[0070] Specifically, through the comprehensive extraction of system logs, security logs, application logs, user access records, operation logs and database contents, the comprehensiveness and integrity of the audit data are ensured, providing a solid foundation for subsequent analysis.

[0071] Further, S2 includes the following steps:

[0072] S201: Data cleaning, ensuring data accuracy and consistency by removing duplicate data, correcting erroneous data, and deleting irrelevant data;

[0073] S202: Missing value processing: filling or deleting missing data using commonly used methods such as mean filling, interpolation, or model-based prediction filling;

[0074] S203: Data standardization: standardize the numerical data to make it conform to a unified scale. Common methods include z-score standardization and min-max normalization.

[0075] Specifically, data cleaning, missing value processing and data standardization were implemented to effectively improve the quality and consistency of the data, laying the foundation for accurate model training. The effectiveness of the data was improved by removing noise and redundant data.

[0076] Furthermore, S3 includes the following steps:

[0077] S301: Feature extraction: extracting information that can represent data features from raw data through feature engineering. Methods used include principal component analysis (PCA), linear discriminant analysis (LDA), etc.

[0078] S302: Feature screening, using correlation analysis, information gain and other methods to screen out the most representative features and remove redundant and noisy features;

[0079] S303: Feature conversion: reconstruct the features through feature conversion techniques such as polynomial conversion and frequency domain transformation to improve the expressiveness of the features and the generalization ability of the model.

[0080] Specifically, through feature extraction, screening and conversion, we ensure that the most representative and informative features are used during model training, improving the model's expressiveness and generalization capabilities. PCA, LDA and other methods are used to further enhance the expressiveness of data features.

[0081] Further, S4 includes the following steps:

[0082] S401: Model architecture design, design the architecture of the multi-layer perceptron (MLP) neural network, including the number of nodes and connection methods of the input layer, multiple hidden layers, and output layer;

[0083] S402: Initialize network parameters, initialize the weights and biases of the neural network, and use methods such as Xavier initialization and He initialization to ensure the stability and convergence speed of network training;

[0084] S403: Activation function selection: select an appropriate activation function for each layer, such as ReLU, Sigmoid, Tanh, etc., to improve the nonlinear expression ability and training efficiency of the model.

[0085] Specifically, we designed and optimized the multi-layer perceptron neural network architecture, and improved the model's ability to perceive complex patterns through multi-layer nonlinear transformations of deep learning. By using appropriate activation functions and network parameter initialization, we ensured the stability and efficiency of model training.

[0086] Furthermore, S5 includes the following steps:

[0087] S501: training set and test set division: divide the preprocessed data into training set and test set, usually using cross-validation or holdout method to ensure fairness in model training and evaluation;

[0088] S502: Hyperparameter adjustment: Tune the model's hyperparameters, such as learning rate, batch size, regularization parameters, etc., through grid search, random search, or Bayesian optimization.

[0089] S503: Gradient descent optimization, using gradient descent algorithms (such as SGD, Adam, etc.) to optimize the parameters of the model, and through multiple iterative training, minimize the loss function and improve the accuracy of the model.

[0090] Specifically, the accuracy and robustness of the model were significantly improved by dividing the training set and test set, tuning the hyperparameters, and using gradient descent optimization. The cross-validation method and advanced hyperparameter optimization methods were used to ensure consistent performance of the model on different data sets.

[0091] Further, S6 includes the following steps:

[0092] S601: Build a data stream architecture and a real-time data stream processing framework (such as Apache Kafka or Flink) to achieve real-time data collection, transmission, and processing;

[0093] S602: Data flow monitoring: real-time monitoring of the status and performance of data flows to ensure the stability and efficiency of data flow processing and timely detection and resolution of abnormal situations;

[0094] S603: Real-time analysis and processing: Integrate real-time analysis algorithms to process and analyze real-time data streams, and promptly detect and respond to potential sensitive data leakage incidents.

[0095] Specifically, through the construction, monitoring and analysis of real-time data flow architecture, the real-time and efficient audit of sensitive data is ensured. It can quickly respond to potential sensitive data leakage incidents and improve the security response speed of the system.

[0096] Furthermore, S7 includes the following steps:

[0097] S701: Evaluation indicator selection: select appropriate evaluation indicators, such as accuracy, recall, F1-score, ROC-AUC, etc., to comprehensively evaluate the model performance;

[0098] S702: Cross-validation: Use the cross-validation method to evaluate the stability and generalization ability of the model through multiple training and validation to prevent overfitting;

[0099] S703: Model tuning: Based on the evaluation results, the model is tuned and improved by adjusting hyperparameters, increasing or decreasing the number of network layers, improving feature engineering, etc., to improve model performance.

[0100] Specifically, by selecting appropriate evaluation indicators and cross-validation, the performance and stability of the model are ensured. Model tuning is performed based on the evaluation results to improve the detection capability and adaptability of the model and prevent overfitting.

[0101] Further, S8 includes the following steps:

[0102] S801: Anomaly detection, by analyzing the output of the model, identifying potential abnormal behaviors and sensitive data leakage events, and performing detailed classification and labeling;

[0103] S802: Security incident report: Generate a security incident report to record in detail the sensitive data leakage incidents found, including time, location, incident description and scope of impact;

[0104] S803: Security measures recommendations: Based on the analysis results, corresponding security measures and improvement suggestions are proposed, such as strengthening authority management, increasing data encryption strength, improving log monitoring, etc.

[0105] Specifically, through anomaly detection and security incident reporting, sensitive data leakage incidents are timely identified and recorded, effective security measures and improvement suggestions are proposed, and the security and defense capabilities of the system are enhanced.

[0106] Furthermore, S9 includes the following steps:

[0107] S901: Differential privacy mechanism design: Introduce differential privacy mechanism in data processing and model training, protect user privacy and prevent sensitive information leakage by adding noise and other methods;

[0108] S902: Privacy protection evaluation: evaluate the effectiveness of the differential privacy mechanism to ensure that it can provide adequate privacy protection while ensuring data availability;

[0109] S903: Privacy policy formulation: formulate and implement privacy protection policies to ensure strict compliance with privacy protection laws, regulations and industry standards during data collection, processing, analysis and storage.

[0110] Specifically, a differential privacy mechanism is introduced during data processing and model training to protect user privacy by adding noise, effectively preventing the leakage of sensitive information. At the same time, through privacy protection assessment and policy formulation, the legality, compliance and security of data use are ensured.

[0111] The present invention provides a sensitive data audit method based on neural network multi-layer perception. First, all relevant data in the target system are comprehensively collected through log file acquisition, user behavior record collection and database content extraction to ensure the comprehensiveness and integrity of the data. Then, the collected data is cleaned, missing value processed and standardized to remove noise and redundant data, and improve the quality and consistency of the data. After data preprocessing, meaningful features are extracted through feature extraction, screening and conversion to ensure that the most representative information is used during model training. Then, a multi-layer perceptron (MLP) neural network architecture is designed and optimized, the perception ability and accuracy of the model are improved through deep learning, and appropriate activation functions and network parameter initialization are used to ensure the stability and efficiency of model training. In the model training stage, the accuracy and robustness of the model are improved by dividing the training set and the test set, adjusting the hyperparameters and using the gradient descent optimization algorithm. Real-time data stream processing realizes the real-time and efficiency of sensitive data auditing by building a data stream architecture, real-time monitoring and analyzing the data stream. In the model evaluation stage, appropriate evaluation indicators and cross-validation methods are selected to ensure model performance and stability, and the model is tuned and improved according to the evaluation results. Finally, through anomaly detection and security incident reporting, sensitive data leakage incidents are promptly identified and recorded, and effective security measures and improvement suggestions are proposed. At the same time, a differential privacy mechanism is introduced in the process of data processing and model training to protect user privacy by adding noise, ensuring the legality, compliance and security of data use. Through the above steps, the sensitive data auditing method based on neural network multi-layer perception effectively improves the accuracy, real-time and security of data auditing.

[0112] The above description is only a preferred embodiment of the present invention and does not limit the present invention in other forms. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A sensitive data auditing method based on neural network multi-layer perception, characterized in that: The following steps are involved: S1: Data collection, collecting all relevant data in the target system, including log files, user behavior records, database content, etc.; S2: Data preprocessing: cleaning and preprocessing the collected data, removing noise data, filling missing values, and standardizing the data; S3: Feature selection, extracting meaningful features from preprocessed data through feature engineering for subsequent model training and analysis; S4: Multi-layer perceptron neural network integration, using multi-layer perceptron (MLP) neural network to conduct deep learning on data, and improve the perception ability and accuracy of the model through multi-layer nonlinear transformation; S5: Model training, using traditional machine learning algorithms (such as logistic regression, decision trees, etc.) to model the data and train the model to identify potential sensitive data leakage or abnormal behavior; S6: Real-time data stream processing, integrating with real-time data stream processing frameworks (such as Apache Kafka or Flink), so that sensitive data audits can be performed in real time instead of relying on offline data analysis; S7: Model evaluation: evaluate the trained model through cross-validation and other evaluation methods to ensure that its performance and accuracy meet the requirements; S8: Result analysis: Analyze the results of the model output, identify and report possible sensitive data leakage incidents, and propose corresponding security measures; S9: Enhanced privacy protection. Differential privacy technology is used in data processing and model training to ensure that user privacy is not leaked during sensitive data audits, thereby improving the security and trust of the system.

2. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized in that: The S1 comprises the following steps: S101: Log file acquisition, extracting all relevant log files from the target system, including system logs, security logs, application logs, etc., to ensure full coverage of system activities; S102: User behavior record collection, by monitoring and recording user operation behaviors, collecting user access records, operation logs, login records, etc., for subsequent behavior analysis; S103: Database content extraction, obtaining all necessary data tables and fields from the database, including structured and unstructured data, to ensure data comprehensiveness and integrity.

3. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized in that: The S2 comprises the following steps: S201: Data cleaning, ensuring data accuracy and consistency by removing duplicate data, correcting erroneous data, and deleting irrelevant data; S202: Missing value processing: filling or deleting missing data using commonly used methods such as mean filling, interpolation, or model-based prediction filling; S203: Data standardization: standardize the numerical data to make it conform to a unified scale. Common methods include z-score standardization and min-max normalization.

4. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized in that: The S3 comprises the following steps: S301: Feature extraction: extracting information that can represent data features from raw data through feature engineering. Methods used include principal component analysis (PCA), linear discriminant analysis (LDA), etc. S302: Feature screening, using correlation analysis, information gain and other methods to screen out the most representative features and remove redundant and noisy features; S303: Feature conversion: reconstruct the features through feature conversion techniques such as polynomial conversion and frequency domain transformation to improve the expressiveness of the features and the generalization ability of the model.

5. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized in that: The S4 comprises the following steps: S401: Model architecture design, design the architecture of the multi-layer perceptron (MLP) neural network, including the number of nodes and connection methods of the input layer, multiple hidden layers, and output layer; S402: Initialize network parameters, initialize the weights and biases of the neural network, and use methods such as Xavier initialization and He initialization to ensure the stability and convergence speed of network training; S403: Activation function selection: select an appropriate activation function for each layer, such as ReLU, Sigmoid, Tanh, etc., to improve the nonlinear expression ability and training efficiency of the model.

6. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized in that: The S5 comprises the following steps: S501: training set and test set division: divide the preprocessed data into training set and test set, usually using cross-validation or holdout method to ensure fairness in model training and evaluation; S502: Hyperparameter adjustment: Tune the model's hyperparameters, such as learning rate, batch size, regularization parameters, etc., through grid search, random search, or Bayesian optimization. S503: Gradient descent optimization, using gradient descent algorithms (such as SGD, Adam, etc.) to optimize the parameters of the model, and through multiple iterative training, minimize the loss function and improve the accuracy of the model.

7. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized in that: The S6 comprises the following steps: S601: Build a data stream architecture and a real-time data stream processing framework (such as Apache Kafka or Flink) to achieve real-time data collection, transmission, and processing; S602: Data flow monitoring: real-time monitoring of the status and performance of data flows to ensure the stability and efficiency of data flow processing and timely detection and resolution of abnormal situations; S603: Real-time analysis and processing: Integrate real-time analysis algorithms to process and analyze real-time data streams, and promptly detect and respond to potential sensitive data leakage incidents.

8. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized by: The S7 comprises the following steps: S701: Evaluation indicator selection: select appropriate evaluation indicators, such as accuracy, recall, F1-score, ROC-AUC, etc., to comprehensively evaluate the model performance; S702: Cross-validation: Use the cross-validation method to evaluate the stability and generalization ability of the model through multiple training and validation to prevent overfitting; S703: Model tuning: Based on the evaluation results, the model is tuned and improved by adjusting hyperparameters, increasing or decreasing the number of network layers, improving feature engineering, etc., to improve model performance.

9. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized in that: The S8 comprises the following steps: S801: Anomaly detection, by analyzing the output of the model, identifying potential abnormal behaviors and sensitive data leakage events, and performing detailed classification and labeling; S802: Security incident report: Generate a security incident report to record in detail the sensitive data leakage incidents found, including time, location, incident description and scope of impact; S803: Security measures recommendations: Based on the analysis results, corresponding security measures and improvement suggestions are proposed, such as strengthening authority management, increasing data encryption strength, improving log monitoring, etc.

10. The sensitive data auditing method based on neural network multi-layer perception according to claim 1 is characterized by: The S9 includes the following steps: S901: Differential privacy mechanism design: Introduce differential privacy mechanism in data processing and model training, protect user privacy by adding noise and other methods, and prevent sensitive information leakage; S902: Privacy protection evaluation: evaluate the effectiveness of the differential privacy mechanism to ensure that it can provide adequate privacy protection while ensuring data availability; S903: Privacy policy formulation: formulate and implement privacy protection policies to ensure strict compliance with privacy protection laws, regulations and industry standards during data collection, processing, analysis and storage.

Images