Dynamic data adaptive desensitization method and device based on artificial intelligence

Abstract

The embodiments of the present application provide a method and device for adaptive desensitization of dynamic data based on artificial intelligence, which relate to the technical field of dynamic data desensitization. The method includes: labeling sensitive information in dynamic data based on deep learning models and natural language processing technology; training a reinforcement learning model to generate a dynamic desensitization strategy for dynamic data to perform data desensitization processing; identifying and classifying the association relationships in dynamic data through a graph neural network, and uniformly processing the associated data; completing the desensitization processing of the dynamic data, receiving desensitization feedback data, and adjusting the dynamic desensitization strategy based on the feedback desensitized data; ensuring the accuracy and comprehensiveness of the desensitization processing; automatically generating and adjusting the desensitization strategy to save workload and avoid omissions; identifying and processing associated data to ensure the integrity and consistency of the data; and ensuring the effectiveness of the desensitized data in subsequent data analysis and machine learning model training.

Description

Technical Field

[0001] The present application relates to the technical field of dynamic data desensitization, and in particular to a method and device for adaptive desensitization of dynamic data based on artificial intelligence. Background Art

[0002] Dynamic data masking is a technology that performs special processing on data during runtime, aiming to mask the true nature of sensitive data to protect data privacy. Unlike static data masking, dynamic data masking is performed not during data storage but during data query or use. This ensures that users do not need to use real data during actual operations, but rather the masked data, effectively avoiding direct exposure of data information.

[0003] Dynamic data masking is implemented through built-in security policies. Users can define masking policies based on their needs, including masking methods and user filters. In practice, when users query tables containing sensitive data, the system maskes the target fields according to the defined masking policy, using methods such as replacement, scrambling, or generating random values ​​to protect data privacy.

[0004] The advantage of dynamic data masking is that it allows you to set different masking policies for different users in a multi-user environment, ensuring that each user can only access the masked data within their permissions. Furthermore, it prevents data leaks and unauthorized access. Even if an attacker gains access to database files, they will only see the masked data. Dynamic data masking can also help enterprises meet compliance requirements and audit needs, ensuring that only masked data is displayed during audits through masking policies.

[0005] Dynamic data desensitization can be applied to a variety of scenarios, however, the current dynamic desensitization technology has the following shortcomings and deficiencies: Inaccurate identification of sensitive data: Current dynamic desensitization technology relies on manual labeling of sensitive data, and the accuracy and comprehensiveness of the desensitization process are difficult to guarantee; Difficulty in selecting desensitization algorithms: In order to implement a sophisticated desensitization strategy, it is necessary to configure the corresponding algorithm in detail for each application, each interface, and each field. This not only requires a large configuration workload, but is also prone to omissions; Invalidation of data association relationships: After dynamic desensitization processing, the original association relationships between data are destroyed, affecting the integrity and consistency of the data; The quality of desensitized data is not high: The statistical characteristics and distribution patterns of the data cannot be effectively retained, and the data cannot be effectively used in subsequent data analysis and machine learning model training. Summary of the Invention

[0006] The purpose of the embodiments of the present application is to provide a dynamic data adaptive desensitization method and device based on artificial intelligence to solve the problems existing in the existing dynamic data desensitization methods.

[0007] In a first aspect, an embodiment of the present application provides a dynamic data adaptive desensitization method based on artificial intelligence, comprising:

[0008] Label sensitive information in dynamic data based on deep learning models and natural language processing technology;

[0009] Train reinforcement learning models to generate dynamic desensitization strategies for dynamic data to perform data desensitization;

[0010] Identify and classify the association relationships in dynamic data through graph neural networks, and process the associated data in a unified manner;

[0011] Complete the desensitization processing of dynamic data, receive desensitization feedback data, and adjust the dynamic desensitization strategy based on the feedback desensitization data.

[0012] In the above implementation process, the embodiment of the present application marks sensitive information in dynamic data based on deep learning models and natural language processing technology; trains reinforcement learning models to generate dynamic desensitization strategies for dynamic data to perform data desensitization processing; identifies and classifies correlation relationships in dynamic data through graph neural networks, and uniformly processes related data; completes desensitization processing of dynamic data, receives desensitization feedback data, and adjusts dynamic desensitization strategies based on the feedback desensitized data; automatically identifies sensitive information in data to ensure the accuracy and comprehensiveness of desensitization processing; uses adaptive desensitization algorithms to automatically generate and adjust desensitization strategies to save workload and avoid omissions; identifies and processes related data to ensure data integrity and consistency; adjusts desensitization strategies through feedback desensitized data to ensure the effectiveness of desensitized data in subsequent data analysis and machine learning model training.

[0013] Furthermore, the labeling of sensitive information in dynamic data based on deep learning models and natural language processing technology includes:

[0014] Determine the labeling target and the type of sensitive data to be identified;

[0015] According to the annotation rules, select the annotation tool and input the training data to train the deep learning model;

[0016] Based on deep learning models combined with natural language processing technology, the data type of dynamic data is identified, and the corresponding annotation method is called to annotate sensitive information.

[0017] In the above implementation process, deep learning models are combined with natural language processing technology to identify and standardize sensitive information in dynamic data; it can accurately identify sensitive data such as personal information and commercial secrets, and at the same time support customized sensitive vocabulary to improve the flexibility and accuracy of recognition.

[0018] Furthermore, the training reinforcement learning model and generating a dynamic desensitization strategy based on dynamic data to perform desensitization processing include:

[0019] Obtain training data and perform preprocessing and context analysis on the data to obtain context information of the data;

[0020] Define the reinforcement learning environment, select the reinforcement learning algorithm to train the reinforcement learning model, and optimize and adjust the model;

[0021] When a data desensitization request is received, the characteristics and context information of the dynamic data are input into the trained reinforcement learning model to generate a desensitization strategy through the reinforcement learning model;

[0022] According to the generated desensitization strategy, call the corresponding desensitization tool or function to desensitize the data.

[0023] In the above implementation process, through reinforcement learning models and dynamic desensitization strategies, the desensitization intensity and method are dynamically adjusted to strike a balance between protecting privacy and maintaining data availability.

[0024] Furthermore, the identification and classification of association relationships in dynamic data through graph neural networks and unified processing of the associated data include:

[0025] Preprocess and format conversion of training data;

[0026] Select and build a graph neural network model;

[0027] Train and optimize the graph neural network model through training data;

[0028] The dynamic data is input into the trained graph neural network model to identify and classify the association relationships in the dynamic data and perform unified processing on the associated data.

[0029] In the above implementation process, the graph neural network analyzes the dependencies and structures of the data, processes the related data in a unified manner, and maintains the integrity and consistency of the data set.

[0030] Furthermore, the acquisition of training data and preprocessing and context analysis of the data to obtain context information of the data includes:

[0031] Collect training data from various data sources, including databases, file systems, and application programming interfaces, and data types including structured data, semi-structured data, and unstructured data;

[0032] If the data type is structured data, obtain the table structure and data content through the database connection;

[0033] If the data type is semi-structured or unstructured data, use the corresponding parsing tools to parse and extract key information;

[0034] Clean the collected training data to remove noise, duplicate values ​​and outliers;

[0035] Extract features from training data based on data type and business requirements;

[0036] Analyze the context information of the training data, where the context information includes: the source of the data, the purpose of the data, the access frequency of the data, and the association relationship of the data;

[0037] Construct a data graph to express the relationship between training data in a set way;

[0038] The steps of defining a reinforcement learning environment, selecting a reinforcement learning algorithm to train a reinforcement learning model, and optimizing and adjusting the model include:

[0039] Define the reinforcement learning environment, where the reinforcement learning environment includes the state space, action space and reward function;

[0040] Define the state space, action space, and reward function;

[0041] Select a reinforcement learning algorithm;

[0042] Adjust the hyperparameters of the reinforcement learning algorithm to obtain the hyperparameter combination for the current task through experimentation and tuning. The hyperparameters include learning rate, discount factor, experience replay buffer size, and batch size.

[0043] Use exploration and exploitation balance strategy to train reinforcement learning models;

[0044] Regularly evaluate and validate the reinforcement learning model, and use an independent test set to test the reinforcement learning model to adjust and optimize the model;

[0045] When a data desensitization request is received, the characteristics and context information of the dynamic data are input into the trained reinforcement learning model to generate a desensitization strategy through the reinforcement learning model, including:

[0046] When a data desensitization request is received, the features and contextual information of the dynamic data are input into the trained reinforcement learning model;

[0047] Based on the reinforcement learning model, the optimal action is selected according to the current state, and the corresponding desensitization operation and parameters are determined to generate a desensitization strategy;

[0048] The method of calling the corresponding desensitization tool or function to perform desensitization processing on the data according to the generated desensitization strategy includes:

[0049] According to the generated desensitization strategy, for different types of data and desensitization operations, call the corresponding desensitization tool or function to desensitize the data;

[0050] Protect the integrity and availability of data when implementing redaction policies;

[0051] Record the execution process and results of the desensitization strategy, where the execution results include the original data, desensitized data, the time of the desensitization operation, and the person who executed it.

[0052] Furthermore, the preprocessing and format conversion of the training data includes:

[0053] Obtain training data from different data sources and preprocess the training data, including removing duplicate data and handling missing values;

[0054] Extracting corresponding features from node data of different data sources; wherein the features include numerical, categorical or textual feature data;

[0055] Standardize or normalize the features so that different features have the same scale and distribution;

[0056] Define the nodes and edges of the graph, where nodes correspond to data entities and edges represent the relationships between entities;

[0057] Construct an adjacency matrix based on the relationship information in the data and create a feature matrix for each node;

[0058] The selecting and constructing of a graph neural network model includes:

[0059] Select the type of graph neural network model based on the characteristics of the training data and task requirements;

[0060] The training and optimization of the graph neural network model using training data includes:

[0061] Receive the feature matrix and adjacency matrix of the graph as input layer;

[0062] Select the corresponding hidden layer according to the type of graph neural network model;

[0063] Design the output layer according to the task type;

[0064] Divide the input layer into training set, validation set and test set;

[0065] Define the loss function and select the optimization algorithm;

[0066] Initialize the model parameters, extract a set amount of data from the training set, input the set amount of data into the graph neural network model, perform forward propagation calculations, and obtain the prediction results of the graph neural network model;

[0067] Based on the predicted results and the true labels, the value of the loss function is calculated, and the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm;

[0068] Update the model parameters using the calculated gradients according to the selected optimization algorithm;

[0069] Monitor metrics during training, including loss and accuracy.

[0070] The dynamic data is input into a trained graph neural network model to identify and classify the association relationships in the dynamic data and perform unified processing on the association data, including:

[0071] Calculate the embedding representation of each node through the trained graph neural network model;

[0072] Identify and classify related data based on node embeddings;

[0073] The associated data after identification and classification are processed uniformly.

[0074] Furthermore, the receiving of desensitization feedback data and adjusting the dynamic desensitization strategy based on the feedback desensitization data include:

[0075] Establish an evaluation indicator system to measure the effectiveness of dynamic desensitization strategies; the evaluation indicators include: data security, data availability, business compliance, and performance indicators;

[0076] Establish user feedback channels to collect business users' experience and opinions on the use of desensitized data;

[0077] Improve and optimize the reinforcement learning model based on evaluation metrics and user feedback;

[0078] Regularly retrain the reinforcement learning model and incorporate new data and feedback information into the training process to continuously optimize the reinforcement learning model and dynamic desensitization strategy.

[0079] In the above implementation process, by monitoring the performance of the desensitized data in actual applications, collecting feedback, and adjusting the desensitization parameters, we ensure that the data meets the privacy protection requirements while enhancing the effectiveness of the desensitized data in subsequent data analysis and machine learning model training.

[0080] In a second aspect, an embodiment of the present application provides a dynamic data adaptive desensitization device based on artificial intelligence, comprising:

[0081] Intelligent sensitive data identification module, used to mark sensitive information in dynamic data based on deep learning models and natural language processing technology;

[0082] Adaptive desensitization algorithm module, used to train reinforcement learning models and generate dynamic desensitization strategies for dynamic data to perform data desensitization processing;

[0083] The data association relationship maintenance module is used to identify and classify the association relationships in dynamic data through graph neural networks and perform unified processing on the associated data;

[0084] The desensitized data quality optimization module is used to complete the desensitization processing of dynamic data, receive desensitization feedback data, and adjust the dynamic desensitization strategy based on the feedback desensitized data.

[0085] In a third aspect, an embodiment of the present application provides an electronic device, including:

[0086] A processor, a memory and a bus, wherein the processor is connected to the memory via the bus, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, they are used to implement the dynamic data adaptive desensitization method based on artificial intelligence as described above.

[0087] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a server, the dynamic data adaptive desensitization method based on artificial intelligence as described above is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0089] Figure 1 A flowchart of a dynamic data adaptive desensitization method based on artificial intelligence provided in an embodiment of the present application;

[0090] Figure 2 This is a structural diagram of an artificial intelligence-based dynamic data adaptive desensitization device provided in an embodiment of the present application;

[0091] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0092] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0093] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0094] Dynamic data desensitization is applicable to a variety of scenarios, including but not limited to: data sharing in a multi-user environment, ensuring that different users can only access data within their authority; preventing data leakage and illegal access, and reducing the risk of unauthorized access to sensitive information; meeting compliance requirements and protecting personal privacy through desensitization; providing desensitized data in big data analysis and development and testing environments to protect the security of real sensitive data; the development and application of dynamic data desensitization technology is of great significance for modern enterprises to protect sensitive data during the digital transformation process.

[0095] Current dynamic desensitization technology has the following shortcomings and deficiencies: Inaccurate identification of sensitive data: Current dynamic desensitization technology relies on manual labeling of sensitive data, and the accuracy and comprehensiveness of the desensitization process are difficult to guarantee; Difficulty in selecting desensitization algorithms: In order to implement a sophisticated desensitization strategy, it is necessary to configure the corresponding algorithm in detail for each application, each interface, and each field. This not only requires a large configuration workload, but is also prone to omissions; Invalidation of data association relationships: After dynamic desensitization processing, the original association relationships between data are destroyed, affecting the integrity and consistency of the data; The quality of desensitized data is not high: The statistical characteristics and distribution patterns of the data cannot be effectively retained, and the data cannot be effectively used in subsequent data analysis and machine learning model training.

[0096] Please see Figure 1 , Figure 1 A flowchart of an artificial intelligence-based dynamic data adaptive desensitization method provided in an embodiment of the present application. The artificial intelligence-based dynamic data adaptive desensitization method includes:

[0097] 100. Based on deep learning models and natural language processing technology, label sensitive information in dynamic data.

[0098] 110. Determine the labeling target and the type of sensitive data to be identified.

[0099] Optionally, clearly label the target and the type of sensitive data to be identified, such as personal identity information (name, ID number, phone number, etc.), financial information (bank account number, credit card number, etc.), medical information (medical records, diagnosis results, etc.), etc.

[0100] Optionally, determine the level of detail and standards for annotation based on specific application scenarios and requirements. For example, you can form an annotation team: select personnel with relevant domain knowledge and annotation experience to form the annotation team. This team may include experts in specialized fields, data analysts, and annotators. Provide training to familiarize the annotation team with the annotation objectives, annotation standards, and how to use annotation tools.

[0101] Understandably, automatic labeling can also be performed through AI automatic labeling technology. Specifically, by training and learning large amounts of data, it can automatically identify and classify data, making it familiar with labeling targets, labeling standards, and how to use labeling tools, thereby achieving automatic labeling of data.

[0102] Specifically, according to the characteristics of the data type and the labeling task, select an appropriate labeling tool, such as LabelImg, BRAT, etc.; optionally, the labeling tool can have convenient labeling operations, data management, multi-person collaboration and other functions.

[0103] 120. According to the annotation rules, select the annotation tool and input the training data to train the deep learning model.

[0104] Optionally, the data to be labeled is divided into a training set, a validation set, and a test set. The training set is used for model training, the validation set is used for model tuning, and the test set is used to evaluate model performance. For example, 70% of the data is used as the training set, 20% as the validation set, and 10% as the test set. The resulting training set is then fed into the model for training.

[0105] 130. Based on the deep learning model combined with natural language processing technology, the data type of dynamic data is identified, and the corresponding annotation method is used to annotate sensitive information.

[0106] It is understandable that different data types of dynamic data are identified and corresponding annotation methods are called to annotate sensitive information.

[0107] For example, for text data: Read the text line by line or paragraph by paragraph. When sensitive data is found, use annotation tools to select the sensitive data in the text and add corresponding labels. For example, annotate a name as "PERSON" and an ID number as "ID_NUMBER." For complex text, contextual analysis and annotation may be required.

[0108] For example, for image data: Check the image content. If the image contains sensitive information, such as faces or license plate numbers in personal photos, you can use annotation tools to draw a box on the image and add corresponding labels. Different types of sensitive information may require different annotation methods, such as polygon annotation or key point annotation.

[0109] For example, for other types of data, such as audio, video, etc., corresponding annotation methods can be adopted according to specific needs.

[0110] Additionally, regular quality checks can be conducted during the annotation process. A random sample of annotated data can be reviewed by other annotators, experts, or AI to ensure accuracy and consistency. Any inconsistencies or disputes can be discussed and negotiated to achieve unified annotation standards. A standard annotation document can be created to document issues and solutions during the annotation process for future reference and improvement.

[0111] In addition, the labeled data should be organized and archived, and the labeled data should be stored separately according to the training set, validation set and test set, and the data format and naming conventions should be unified; optionally, the labeled data can be converted into the format required by the deep learning model, such as CSV, JSON, TFRecord, etc.; optionally, a final review of the organized labeled data should be conducted to check whether the data is complete, the labels are accurate, the format is correct, etc. If any problems are found, they can be corrected and improved in a timely manner.

[0112] As described above, the embodiments of the present application utilize a deep learning model combined with natural language processing technology to identify and standardize sensitive information in dynamic data; it can accurately identify sensitive data such as personal information and commercial secrets, while supporting customized sensitive word libraries to improve the flexibility and accuracy of recognition.

[0113] 200. Train the reinforcement learning model to generate a dynamic desensitization strategy for dynamic data to perform data desensitization.

[0114] 210. Obtain training data and perform preprocessing and context analysis on the data to obtain context information of the data.

[0115] Specifically, training data is collected from various data sources, including databases, file systems, and application programming interfaces (APIs). Data types include structured data (such as tables in relational databases), semi-structured data (such as JSON and XML files), and unstructured data (such as text files and images). If the data type is structured, the table structure and data content are obtained through a database connection. If the data type is semi-structured or unstructured, the corresponding parsing tools are used to parse and extract key information.

[0116] Specifically, training data is preprocessed: collected training data is cleaned to remove noise, duplicate values, and outliers. For example, for text data, natural language processing techniques are used to perform preprocessing operations such as lexical analysis and syntactic analysis to remove irrelevant information such as stop words and punctuation. Furthermore, feature extraction is performed on the training data based on the data type and business requirements. For example, for image data, color features, texture features, shape features, etc. can be extracted; for structured data, features such as field name, data type, data length, and primary and foreign key relationships are extracted.

[0117] Specifically, contextual analysis is performed on the training data: analyzing the contextual information of the training data, including the data's source, purpose, access frequency, and relationships. For example, database logs are used to analyze data access patterns to understand which users accessed which data at what time. Business processes are used to analyze the flow and usage of data in different business processes. Furthermore, a data graph is constructed to represent the relationships between the training data in a predefined manner, which can be graphically represented. For example, for a relational database, an entity relationship diagram can be constructed to display the relationships between tables.

[0118] 220. Define the reinforcement learning environment, select the reinforcement learning algorithm to train the reinforcement learning model, and optimize and adjust the model.

[0119] Optionally, define a reinforcement learning environment, where the reinforcement learning environment includes a state space, an action space, and a reward function.

[0120] Specifically, the state space consists of data features and contextual information, for example, including data type, sensitivity level, access frequency, business scenario and other features. The dimension of the state space is determined according to actual conditions, and each dimension represents a feature.

[0121] Specifically, the action space is used to define possible desensitizing operations. For example, for numerical data, there may be operations such as truncation, rounding, and blurring; for text data, there may be operations such as replacement, encryption, and anonymization; for image data, there may be operations such as blurring and pixelation; the size of the action space is determined according to the specific type of desensitizing operation.

[0122] Specifically, the reward function defines rewards based on the effectiveness of the desensitization strategy and business needs. The goal of the reward function is to encourage the generation of desensitization strategies that both meet business needs and effectively protect data security. For example, if the desensitized data can minimize the risk of sensitive information leakage while meeting business needs, a higher reward will be given; conversely, if the desensitization strategy causes business operations to fail or severely reduces data availability, a lower reward will be given.

[0123] Optionally, select and train the model; specifically, choose reinforcement learning algorithms such as Deep Q-Network (DQN) and Deep Deterministic Policy Gradient (DDPG). These algorithms can process continuous state space and action space and are suitable for complex desensitization strategy generation tasks.

[0124] For example, for the DQN algorithm: First, a neural network is constructed as an approximator of the Q-value function, usually using a multi-layer convolutional neural network (CNN) or a multi-layer perceptron (MLP). The input of the network is the state vector, and the output is the Q value corresponding to each action.

[0125] Secondly, through the experience replay mechanism, the experience (state, action, reward, next state) generated by the interaction between the agent and the environment is stored in a replay buffer. During training, small batches of experience data are randomly extracted from the replay buffer for training to break the correlation between the data and improve training stability.

[0126] In addition, the target network is used to stabilize the training process. The parameters of the current Q network are regularly copied to the target network, which is used to calculate the target Q value, that is, the target Q value:

[0127] ;

[0128] in, is the current reward, is the discount factor, is the next state, It's the next move. are the parameters of the target network, It means that the current Q network is in state Next for action The predicted Q value of .

[0129] The parameters of the current Q network are updated by minimizing the loss function, which is:

[0130] ;

[0131] in, and is the current state and action, are the parameters of the current Q network.

[0132] For example, for the DQN algorithm: First, construct the actor network and the critic network respectively. The actor network is used to generate actions based on the current state, that is, to generate actions. The actor network is:

[0133] ;

[0134] in, is the policy function of the actor network, are the parameters of the actor network, It's a status.

[0135] The critic network is used to evaluate the value of the current state and action, namely: ,in, are the parameters of the critic network.

[0136] Secondly, the experience replay mechanism and soft update strategy are used for training. The experience replay mechanism is similar to DQN, which is used to store and extract experience data; the soft update strategy refers to slowly updating the parameters of the target network, for example, the target parameters of the actor network and the target parameters of the critic network ,use and Update in the following way, is a small update factor.

[0137] In addition, the training goal of the actor network is to maximize the expected reward, that is, to update the parameters through the gradient ascent algorithm So that:

[0138] ;

[0139] in, is the objective function, and D is the experience replay buffer. The training objective of the critic network is to minimize the mean squared error loss function.

[0140] Optionally, the model is optimized and adjusted. First, the hyperparameters of the reinforcement learning algorithm are adjusted. Through experiments and tuning, a hyperparameter combination for the current task is obtained to improve the training efficiency and performance of the model. The hyperparameters include the learning rate, discount factor, experience replay buffer size, and batch size.

[0141] Secondly, a balanced exploration and exploitation strategy is used to train reinforcement learning models, such as the greedy strategy (in DQN) or the noise-based exploration strategy (in DDPG). In the early stages of training, the agent conducts more exploration, trying different desensitization operations to fully understand the environment. As training progresses, it gradually leverages its existing experience to select the optimal actions, thereby improving the quality of the generated strategy.

[0142] In addition, reinforcement learning models should be regularly evaluated and validated, using independent test sets to test them and adjust and optimize them. If the model performs poorly on the test set, analyze the reasons and make appropriate adjustments, such as adding training data, improving the model structure, and adjusting hyperparameters.

[0143] 230. When a data desensitization request is received, the characteristics and context information of the dynamic data are input into the trained reinforcement learning model to generate a desensitization strategy through the reinforcement learning model.

[0144] Specifically, when a data desensitization request is received, the characteristics and contextual information of the dynamic data are input into the trained reinforcement learning model; based on the reinforcement learning model, the optimal action is selected according to the current state, and the corresponding desensitization operation and parameters are determined to generate a desensitization strategy.

[0145] For example, for text data containing a personal ID number, the model selects an appropriate desensitization operation based on contextual information such as the data type (text), sensitivity level (high, the ID number is sensitive information), and business scenario (which may be data analysis, but does not require the exact ID number). For example, the model replaces the middle digits of the ID number with asterisks ("****").

[0146] The generated desensitization policy is expressed in an executable form, such as a script, a configuration file, or an API call instruction, so that it can be applied in the actual data processing process.

[0147] 240. According to the generated desensitization strategy, call the corresponding desensitization tool or function to desensitize the data.

[0148] Specifically, based on the generated desensitization strategy, for different types of data and desensitization operations, the corresponding desensitization tools or functions are called to perform desensitization on the data. This requires the use of appropriate technologies. For example, string processing functions can be used to replace text data; mathematical functions and data processing libraries can be used to truncate or obfuscate numerical data; and image processing libraries can be used to obfuscate or pixelate image data.

[0149] Specifically, when implementing the desensitization strategy, the integrity and availability of the data must be protected; for example, when desensitizing text data, the semantic coherence and readability of the text must be maintained; when desensitizing image data, the basic features and information of the image must not be completely destroyed to meet certain business needs for data analysis and use.

[0150] Specifically, the execution process and results of the desensitization strategy are recorded, where the execution results include the original data, desensitized data, the time of the desensitization operation and the person who executed it, so as to facilitate auditing and tracing.

[0151] As described above, the embodiments of the present application dynamically adjust the desensitization intensity and method through reinforcement learning models and dynamic desensitization strategies to achieve a balance between protecting privacy and maintaining data availability.

[0152] 300. Use graph neural networks to identify and classify the association relationships in dynamic data and uniformly process the associated data.

[0153] It is understandable that Graph Neural Network (GNN), as a powerful tool, can effectively process graph structured data and mine the associated information in the data.

[0154] 310. Preprocess and format convert the training data.

[0155] Specifically, training data is obtained from different data sources and preprocessed. This includes removing duplicate data and handling missing values. Preprocessing also includes data cleaning, feature extraction, and graph construction, converting the raw data into a graph structure suitable for GNN processing, where nodes represent data entities and edges represent relationships between entities.

[0156] Specifically, for node data from different data sources, corresponding features are extracted; wherein the features include numerical, categorical or textual feature data; illustratively, for user nodes, features such as age, gender, and occupation are extracted; for protein nodes, amino acid sequence features, functional categories, etc. are extracted.

[0157] Specifically, the features are standardized or normalized so that different features have the same scale and distribution, which is conducive to model training and convergence. For example, common normalization methods such as Z-score normalization and Min-Max normalization can be used.

[0158] Specifically, the nodes and edges of the graph are defined, where nodes correspond to data entities and edges represent the relationships between entities; relationships can be of various types, such as friend relationships in social networks, protein interaction relationships in biological networks, etc.

[0159] Specifically, an adjacency matrix is ​​constructed based on the relationship information in the data, and a feature matrix is ​​created for each node. For example, an adjacency matrix A is constructed based on the relationship information in the data, where Indicates whether there is an edge between node i and node j and the weight of the edge (if any); if there is an edge, The value of is the weight of the edge; if there is no edge, The value of is 0 (for undirected graphs, it is a symmetric matrix).

[0160] At the same time, a feature matrix X is created for each node, where each row of the feature matrix represents the feature vector of a node. The number of rows in the feature matrix is ​​the same as the number of nodes, and the number of columns is the same as the dimension of the extracted features.

[0161] 320. Select and build a graph neural network model.

[0162] Specifically, the type of graph neural network model is selected based on the characteristics of the training data and task requirements.

[0163] Optionally, select a suitable GNN model architecture based on data characteristics and task requirements, such as Graph Convolutional Network (GCN), Graph Attention Network (GAT), Gated Graph Neural Network (GGNN), etc., and define hyperparameters such as the number of model layers, node feature dimensions, and hidden layer size. Construct the model's input layer, hidden layer, and output layer, and design a suitable neural network structure and activation function.

[0164] For example, GCN is suitable for processing data with regular graph structures. It propagates information on the graph through convolution operations, can effectively capture the local structural information of nodes, and performs well for tasks such as node classification and graph classification.

[0165] GAT introduces an attention mechanism that can adaptively assign different weights to different neighbor nodes, focusing more on the information of important neighbor nodes. It works well for processing graph data with neighbors of different importance, such as analyzing the relationship between users and products in recommendation systems.

[0166] GGNN uses a gating mechanism to control the propagation and update of information in the graph. It is suitable for processing dynamic graphs or graph data with long-range dependencies. For example, it is widely used in processing time series-related graph data.

[0167] It is understandable that the appropriate model should be selected based on the characteristics of the specific data and the task requirements. For example, if the graph structure of the data is relatively regular and focuses on the local neighborhood information of the node, GCN may be a better choice; if the importance of neighboring nodes varies greatly, GAT may be more appropriate; if the data is dynamic or has long-range dependencies, GGNN is worth considering.

[0168] 330. Train and optimize the graph neural network model through training data.

[0169] Specifically, the model receives the graph's feature matrix and adjacency matrix as input. For example, the feature matrix X can be directly input into the model, or it can be subjected to some simple linear transformations before input. The adjacency matrix A is used to define the graph's structural information and is used to propagate node information in the model.

[0170] Specifically, select the corresponding hidden layer according to the type of graph neural network model.

[0171] For example, for a GCN-type model, the hidden layer usually contains multiple graph convolution layers. In each graph convolution layer, the feature representation of the node is updated through the convolution operation. For example, using the formula:

[0172] ;

[0173] in, It is The node feature matrix of the layer, is the normalized adjacency matrix, yes The degree matrix of It is The learnable weight matrix of the layer, is the activation function (such as ReLU).

[0174] For GAT-type models, the hidden layer contains multiple graph attention layers. In each graph attention layer, the attention coefficient between nodes is calculated, and then the node features are updated by weighted summing of the features of neighboring nodes according to the attention coefficient. For example, the attention coefficient is calculated using the formula:

[0175] ;

[0176] in, is the attention function, is a learnable weight matrix, and are the feature vectors of node i and node j respectively.

[0177] Then normalize the attention coefficient:

[0178] ;

[0179] in, is the set of neighbor nodes of node i.

[0180] Finally update the characteristics of node i:

[0181] .

[0182] For GGNN-type models, the hidden layer usually contains a gating mechanism such as a gated recurrent unit (GRU) or a long short-term memory network (LSTM). At each time step (corresponding to the information propagation step in the figure), the hidden state of the node is updated based on the current node features, neighbor node features, and the hidden state of the previous moment. For example, the formula is used:

[0183] ;

[0184] ;

[0185] ;

[0186] ;

[0187] in, and They are update gate and reset gate, is a candidate hidden state, , , is a learnable weight matrix, is the current input feature, is the hidden state at the current moment, It is the hidden state at the previous moment.

[0188] Specifically, the output layer is designed based on the task type. For node classification tasks, the output layer is usually a fully connected layer with the number of nodes equal to the number of categories. The softmax function is used to map the node features to the probability distribution of belonging to each category. For example, using the formula:

[0189] ;

[0190] in, is the node feature matrix of the last hidden layer, b is the bias of the fully connected layer, is the probability vector of the node belonging to each category.

[0191] For graph classification tasks, we usually first perform pooling operations on node features (such as global average pooling, global maximum pooling, etc.) to obtain graph-level feature representations, which are then input into the fully connected layer for classification. For example, using the formula:

[0192] ;

[0193] in, It is a graph-level feature, pooling is a pooling function, and then through the formula:

[0194] ;

[0195] Perform graph classification.

[0196] Specifically, the input layer is divided into training set, validation set and test set. For example, 70% of the data is used as the training set, 20% of the data is used as the validation set, and 10% of the data is used as the test set. When dividing, it is necessary to ensure that the distribution of the data in each subset is similar to ensure the generalization ability of the model.

[0197] Specifically, define the loss function and select the optimization algorithm.

[0198] For example, for node classification tasks, a commonly used loss function is the cross-entropy loss function, which measures the difference between the probability distribution predicted by the model and the true label distribution. The formula is:

[0199] ;

[0200] Where N is the number of nodes, C is the number of categories, is the true label of node i belonging to category c (if node i belongs to category c, then ,otherwise ), is the probability that the model predicts that node i belongs to category c.

[0201] For graph classification tasks, you can use the cross entropy loss function or other loss functions suitable for graph-level classification, such as the mean squared error loss function (if the graph label is a continuous value).

[0202] For example, the Adam optimization algorithm is selected. Adam is a commonly used adaptive learning rate optimization algorithm that combines the advantages of the momentum method and RMSProp (root mean square propagation). It can automatically adjust the learning rate and adaptively use different learning rates to update the step size for different parameters. During the training process, the Adam algorithm calculates the first-order moment estimate (i.e., the exponential moving average of the gradient) and the second-order moment estimate (i.e., the exponential moving average of the square of the gradient) of each parameter, and then updates the parameters based on these estimates. The formula is:

[0203] ;

[0204] in, is the parameter value at step t, is the learning rate, is the bias-corrected value of the first-order moment estimate, is the bias-corrected value of the second-order moment estimate, is a small constant (usually , used to prevent the denominator from being 0).

[0205] For example, the SGD optimization algorithm is selected. The SGD stochastic gradient descent algorithm is a simple and commonly used optimization algorithm. In each iteration, it calculates the gradient according to the current mini-batch data and updates the parameters in the opposite direction of the gradient. The formula is:

[0206] ;

[0207] in, In the parameter The gradient of the loss function at is the learning rate; in practical applications, the appropriate optimization algorithm can be selected according to the characteristics of the data tank and the performance of the model.

[0208] Specifically, the model parameters are initialized, optionally using random initialization. In each training round, a set amount (a small batch) of data is extracted from the training set (either a set of nodes or a set of graphs, depending on the data and task). This set amount of data is fed into the graph neural network model, and forward propagation is performed to obtain the model's predictions. Based on the predictions and the true labels, the loss function is calculated, and the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm. The calculated gradients are then used to update the model parameters according to the selected optimization algorithm.

[0209] Specifically, monitor metrics during training, including loss and accuracy. For example, you can regularly evaluate model performance on a validation set to adjust the learning rate, stop training, and so on. For example, if performance on the validation set doesn't improve over several consecutive rounds, consider reducing the learning rate or stopping training early to prevent overfitting.

[0210] 340. Input dynamic data into the trained graph neural network model to identify and classify the association relationships in the dynamic data and uniformly process the associated data.

[0211] Specifically, the embedded representation of each node is calculated through the trained graph neural network model.

[0212] For example, the trained GNN model is used to perform forward propagation on the graph data to calculate the embedding representation of each node. The embedding of the node contains a comprehensive representation of its own characteristics and the information of neighboring nodes, reflecting the dependency and structural characteristics of the data in the graph.

[0213] For the GCN model, the embedding of the node is obtained by calculating multiple graph convolutional layers. After the output of the last graph convolutional layer, each node has a corresponding feature vector, which is the embedding representation of the node.

[0214] For the GAT model, the embedding of a node is obtained by weighted summing of neighbor node features through the graph attention layer, and the node feature vector output by the last graph attention layer is used as the embedding of the node.

[0215] For the GGNN model, the hidden state of the node is output at the last time step as the embedding representation of the node.

[0216] Specifically, related data is identified and classified based on node embedding.

[0217] For example, a clustering algorithm can be used to group nodes with similar embeddings into a single category, thereby discovering related data groups. For example, using the K-Means clustering algorithm, the number of clusters is first determined. Then, the node embeddings are used as input, the distance from each node to the cluster center is calculated, and the node is assigned to the cluster with the closest distance. The cluster centers and node assignments are continuously updated through iteration until the clustering results converge.

[0218] For example, distance-based methods, such as calculating the cosine similarity or Euclidean distance between node embeddings, can also be used to consider nodes with high similarity or close distance as related data. For example, if the cosine similarity between two node embeddings is greater than a certain threshold, they are considered to be related.

[0219] For example, a classifier can be trained to classify nodes using node embeddings as input to identify different types of associated data. For example, a classification algorithm such as a support vector machine (SVM) or decision tree can be used to learn the mapping relationship between node embeddings and category labels on a training set, and then classify the nodes on a test set.

[0220] Specifically, the associated data after identification and classification are processed uniformly.

[0221] For example, the identified and classified related data are processed uniformly. According to specific business needs, the same processing rules can be applied to data of the same category.

[0222] For example, in a data cleaning task, if a set of related data is found to have the same type of noise or outliers, the same method can be used to clean these data, such as deleting outliers, filling missing values, etc.

[0223] In data analysis tasks, related data can be jointly analyzed or modeled. For example, for a set of user behavior data with related relationships, a comprehensive user behavior model can be built to better understand user behavior patterns and preferences.

[0224] In terms of data storage and management, related data can be grouped together for easier query and retrieval. For example, related data belonging to the same business process or project can be stored in a database table or file, and corresponding indexes and associations can be established to facilitate subsequent data processing and analysis.

[0225] As described above, the embodiment of the present application uses a neural network to analyze the dependencies and structures of data, and performs unified processing on the associated data to maintain the integrity and consistency of the data set.

[0226] 400. Complete desensitization processing of dynamic data, receive desensitization feedback data, and adjust the dynamic desensitization strategy based on the feedback desensitization data.

[0227] 410. Establish an evaluation indicator system to measure the effectiveness of dynamic desensitization strategies; the evaluation indicators include: data security, data availability, business compliance and performance indicators.

[0228] Data security is measured by assessing the risk of sensitive information leakage in desensitized data. Methods such as information entropy and differential privacy metrics can be used to quantify data security. For example, the entropy value of sensitive information in desensitized data can be calculated. A larger entropy value indicates increased uncertainty in the sensitive information and improved security.

[0229] Data availability is used to assess the usability of desensitized data for business applications. This can be measured by calculating metrics such as data completeness, accuracy, and consistency. For example, for data analysis tasks, the statistical analysis results of the data before and after desensitization can be compared. If the difference is small, it indicates high data availability.

[0230] Business compliance involves checking whether the desensitization strategy complies with relevant laws, regulations, and business specifications. For example, the desensitization of medical data must ensure compliance with medical privacy regulations; the desensitization of financial data must comply with financial regulatory requirements.

[0231] Performance indicators are used to evaluate the execution efficiency and resource consumption of the desensitization strategy, including desensitization processing time, memory usage, CPU usage, etc.

[0232] 420. Establish user feedback channels to collect business users’ experience and opinions on the use of anonymized data.

[0233] Specifically, establish user feedback channels to collect business users' experience and opinions on the use of desensitized data. Users can provide feedback through interface feedback, questionnaires, online comments, etc. For example, when using desensitized data for data analysis, if business users find that the data has usability issues or does not meet business requirements, they can provide timely feedback to the system.

[0234] 430. Improve and optimize the reinforcement learning model based on evaluation indicators and user feedback.

[0235] Among them, if it is found that the desensitization strategy lacks data security in some cases, the reward function can be adjusted to increase the reward weight for security performance, prompting the model to generate safer strategies; if data availability is greatly affected, the model's training data and feature selection can be adjusted to better consider business needs; if users feedback that certain types of data desensitization methods are unreasonable, the model's action space or strategy generation logic can be improved in a targeted manner.

[0236] 440. Regularly retrain the reinforcement learning model and incorporate new data and feedback information into the training process to continuously optimize the reinforcement learning model and dynamic desensitization strategy.

[0237] As a result, the reinforcement learning model can generate the best desensitization strategy that better meets actual business needs.

[0238] As described above, the embodiments of the present application monitor the performance of the desensitized data in actual applications, collect feedback, and adjust the desensitization parameters to ensure that the data meets the privacy protection requirements while enhancing the effectiveness of the desensitized data in subsequent data analysis and machine learning model training.

[0239] The embodiments of the present application are based on deep learning models and natural language processing technology to mark sensitive information in dynamic data; train reinforcement learning models to generate dynamic desensitization strategies for dynamic data to perform data desensitization processing; identify and classify correlation relationships in dynamic data through graph neural networks, and uniformly process the associated data; complete the desensitization processing of dynamic data, receive desensitization feedback data, and adjust the dynamic desensitization strategy based on the feedback desensitized data; automatically identify sensitive information in the data to ensure the accuracy and comprehensiveness of the desensitization processing; adaptive desensitization algorithms automatically generate and adjust desensitization strategies to save workload and avoid omissions; identify and process associated data to ensure data integrity and consistency; adjust desensitization strategies based on feedback desensitized data to ensure the effectiveness of desensitized data in subsequent data analysis and machine learning model training.

[0240] The above steps are not to be performed in a strict order as described in the numbers, but should be understood as an overall solution.

[0241] In the second aspect, based on the above embodiments, Figure 2 This is a schematic diagram of the structure of a dynamic data adaptive desensitization device based on artificial intelligence provided in an embodiment of the present application. Figure 2 The dynamic data adaptive desensitization device based on artificial intelligence provided in this embodiment specifically includes: an intelligent sensitive data identification module 201, an adaptive desensitization algorithm module 202, a data association relationship maintenance module 203 and a desensitized data quality optimization module 204.

[0242] Among them, the intelligent sensitive data identification module 201 is used to mark sensitive information in dynamic data based on deep learning models and natural language processing technology; the adaptive desensitization algorithm module 202 is used to train the reinforcement learning model to generate a dynamic desensitization strategy for dynamic data to perform data desensitization processing; the data association relationship maintenance module 203 is used to identify and classify the association relationships in dynamic data through graph neural networks, and to uniformly process the associated data; the desensitized data quality optimization module 204 is used to complete the desensitization processing of dynamic data, receive desensitization feedback data, and adjust the dynamic desensitization strategy based on the feedback desensitized data.

[0243] As described above, the embodiments of the present application are based on deep learning models and natural language processing technology to mark sensitive information in dynamic data; train reinforcement learning models to generate dynamic desensitization strategies for dynamic data to perform data desensitization processing; identify and classify the association relationships in dynamic data through graph neural networks, and uniformly process the associated data; complete the desensitization processing of dynamic data, and receive desensitization feedback data, and adjust the dynamic desensitization strategy based on the feedback desensitized data; automatically identify sensitive information in the data to ensure the accuracy and comprehensiveness of the desensitization processing; adaptive desensitization algorithm, automatically generate and adjust the desensitization strategy, save workload, and avoid omissions; identify and process associated data to ensure the integrity and consistency of the data; adjust the desensitization strategy through feedback desensitized data to ensure the effectiveness of the desensitized data in subsequent data analysis and machine learning model training.

[0244] The artificial intelligence-based dynamic data adaptive desensitization device provided in the embodiment of the present application can be used to execute the artificial intelligence-based dynamic data adaptive desensitization method provided in the above embodiment, and has corresponding functions and beneficial effects.

[0245] On the third aspect, an embodiment of the present application also provides an electronic device that can integrate the artificial intelligence-based dynamic data adaptive desensitization device provided in an embodiment of the present application. Figure 3 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Figure 3The electronic device includes: an input device 33, an output device 34, a memory 32, and one or more processors 31; the memory 32 is used to store one or more programs; when the one or more programs are executed by the one or more processors 31, the one or more processors 31 implement the dynamic data adaptive desensitization method based on artificial intelligence as provided in the above embodiment. The input device 33, the output device 34, the memory 32, and the processor 31 can be connected by a bus or other means. Figure 3 The bus connection is taken as an example.

[0246] The processor 31 executes various functional applications and data processing of the device by running software programs, instructions and modules stored in the memory 32, thereby realizing the above-mentioned dynamic data adaptive desensitization method based on artificial intelligence.

[0247] The electronic device provided above can be used to execute the dynamic data adaptive desensitization method based on artificial intelligence provided in the above embodiment, and has corresponding functions and beneficial effects.

[0248] In a fourth aspect, an embodiment of the present application also provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the dynamic data adaptive desensitization method based on artificial intelligence as described above, and can achieve the same beneficial effects.

[0249] Of course, the storage medium containing computer-executable instructions provided in an embodiment of the present application, whose computer-executable instructions are not limited to the dynamic data adaptive desensitization method based on artificial intelligence as described above, can also execute related operations in the dynamic data adaptive desensitization method based on artificial intelligence provided in any embodiment of the present application.

[0250] In a fifth aspect, the embodiments of the present application further provide a computer program product. The methods described in the various embodiments of the present application can be implemented in whole or in part through software, hardware, firmware, or any combination thereof. When implemented using software, they can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the processes or functions described in the various embodiments of the present application are executed in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user device, a core network device, an OAM (Open Application Model), or other programmable device.

[0251] The computer program or instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer program or instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless method. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a digital video disk; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include both volatile and non-volatile types of storage media.

[0252] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0253] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0254] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling an electronic device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0255] The foregoing is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included within the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures.

[0256] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0257] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

Claims

1. A dynamic data adaptive desensitization method based on artificial intelligence, characterized in that: The method comprises: Label sensitive information in dynamic data based on deep learning models and natural language processing technology; Train the reinforcement learning model and generate a dynamic desensitization strategy for dynamic data to perform data desensitization processing; wherein, obtain training data to train the model, define the reinforcement learning environment, and select the reinforcement learning algorithm to train the reinforcement learning model; Among them, training data is collected from various data sources; If the data type is structured data, obtain the table structure and data content through the database connection; If the data type is semi-structured or unstructured data, use the corresponding parsing tools to parse and extract key information; Extract features from training data based on data type and business requirements. For image data, extract color features, texture features, and shape features. For structured data, extract features such as field name, data type, data length, and primary key and foreign key relationships. Among them, the environment of reinforcement learning is defined, and the environment of reinforcement learning includes state space, action space and reward function; The state space consists of data features and context information; The action space is used to define desensitization operations, including truncation, rounding, and blurring for numerical data, replacement, encryption, and anonymization for text data, and blurring and pixelation for image data. The size of the action space is determined by the specific desensitization operation type. The reward function defines the reward based on the effectiveness of the desensitization strategy and business needs; Identify and classify the associations in dynamic data through graph neural networks, and process the associated data in a unified manner; store the associated data in a database table or file, and establish corresponding indexes and associations; Complete the desensitization processing of dynamic data, receive desensitization feedback data, and adjust the dynamic desensitization strategy based on the feedback desensitization data.

2. The method for adaptive desensitization of dynamic data based on artificial intelligence according to claim 1, characterized in that: The method of labeling sensitive information in dynamic data based on deep learning models and natural language processing technology includes: Determine the labeling target and the type of sensitive data to be identified; According to the annotation rules, select the annotation tool and input the training data to train the deep learning model; Based on deep learning models combined with natural language processing technology, the data type of dynamic data is identified, and the corresponding annotation method is called to annotate sensitive information.

3. The method for adaptive desensitization of dynamic data based on artificial intelligence according to claim 1, characterized in that: The training of the reinforcement learning model and the generation of a dynamic desensitization strategy based on dynamic data for desensitization processing include: Obtain training data and perform preprocessing and context analysis on the data to obtain context information of the data; Define the reinforcement learning environment, select the reinforcement learning algorithm to train the reinforcement learning model, and optimize and adjust the model; When a data desensitization request is received, the characteristics and context information of the dynamic data are input into the trained reinforcement learning model to generate a desensitization strategy through the reinforcement learning model; According to the generated desensitization strategy, call the corresponding desensitization tool or function to desensitize the data.

4. The method for adaptive desensitization of dynamic data based on artificial intelligence according to claim 1, characterized in that: The method of identifying and classifying the association relationships in dynamic data through a graph neural network and uniformly processing the association data includes: Preprocess and format conversion of training data; Select and build a graph neural network model; Train and optimize the graph neural network model through training data; The dynamic data is input into the trained graph neural network model to identify and classify the association relationships in the dynamic data and perform unified processing on the associated data.

5. The method for adaptive desensitization of dynamic data based on artificial intelligence according to claim 3 is characterized in that: The acquisition of training data and preprocessing and context analysis of the data to obtain context information of the data includes: Collecting training data from various data sources, wherein the data sources include: databases, file systems, and application programming interfaces, wherein the data types of the training data include: structured data, semi-structured data, and unstructured data; If the data type is structured data, obtain the table structure and data content through the database connection; If the data type is semi-structured or unstructured data, use the corresponding parsing tools to parse and extract key information; Clean the collected training data to remove noise, duplicate values ​​and outliers; Extract features from training data based on data type and business requirements; Analyze the context information of the training data, where the context information includes: the source of the data, the purpose of the data, the access frequency of the data, and the association relationship of the data; Construct a data graph to express the relationship between training data in a set way; The steps of defining a reinforcement learning environment, selecting a reinforcement learning algorithm to train a reinforcement learning model, and optimizing and adjusting the model include: Define the reinforcement learning environment, where the reinforcement learning environment includes the state space, action space and reward function; Define the state space, action space, and reward function; Select a reinforcement learning algorithm; Adjust the hyperparameters of the reinforcement learning algorithm to obtain the hyperparameter combination for the current task through experimentation and tuning. The hyperparameters include learning rate, discount factor, experience replay buffer size, and batch size. Use exploration and exploitation balance strategy to train reinforcement learning models; Regularly evaluate and validate the reinforcement learning model, and use an independent test set to test the reinforcement learning model to adjust and optimize the model; When a data desensitization request is received, the characteristics and context information of the dynamic data are input into the trained reinforcement learning model to generate a desensitization strategy through the reinforcement learning model, including: When a data desensitization request is received, the features and contextual information of the dynamic data are input into the trained reinforcement learning model; Based on the reinforcement learning model, the optimal action is selected according to the current state, and the corresponding desensitization operation and parameters are determined to generate a desensitization strategy; The method of calling the corresponding desensitization tool or function to perform desensitization processing on the data according to the generated desensitization strategy includes: According to the generated desensitization strategy, for different types of data and desensitization operations, call the corresponding desensitization tool or function to desensitize the data; Protect the integrity and availability of data when implementing redaction policies; Record the execution process and results of the desensitization strategy, where the execution results include the original data, desensitized data, the time of the desensitization operation, and the person who executed it.

6. The method for adaptive desensitization of dynamic data based on artificial intelligence according to claim 4 is characterized in that: The preprocessing and format conversion of the training data includes: Obtain training data from different data sources and preprocess the training data, including removing duplicate data and handling missing values; Extracting corresponding features from node data of different data sources; wherein the features include numerical, categorical or textual feature data; Standardize or normalize the features so that different features have the same scale and distribution; Define the nodes and edges of the graph, where nodes correspond to data entities and edges represent the relationships between entities; Construct an adjacency matrix based on the relationship information in the data and create a feature matrix for each node; The selecting and constructing of a graph neural network model includes: Select the type of graph neural network model based on the characteristics of the training data and task requirements; The training and optimization of the graph neural network model using training data includes: Receive the feature matrix and adjacency matrix of the graph as input layer; Select the corresponding hidden layer according to the type of graph neural network model; Design the output layer according to the task type; Divide the input layer into training set, validation set and test set; Define the loss function and select the optimization algorithm; Initialize the model parameters, extract a set amount of data from the training set, input the set amount of data into the graph neural network model, perform forward propagation calculations, and obtain the prediction results of the graph neural network model; Based on the predicted results and the true labels, the value of the loss function is calculated, and the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm; Update the model parameters using the calculated gradients according to the selected optimization algorithm; Monitor metrics during training, including loss and accuracy. The dynamic data is input into a trained graph neural network model to identify and classify the association relationships in the dynamic data and perform unified processing on the association data, including: Calculate the embedding representation of each node through the trained graph neural network model; Identify and classify related data based on node embeddings; The associated data after identification and classification are processed uniformly.

7. The method for adaptive desensitization of dynamic data based on artificial intelligence according to claim 1, characterized in that: The receiving of desensitization feedback data and adjusting the dynamic desensitization strategy based on the feedback desensitization data includes: Establish an evaluation indicator system to measure the effectiveness of dynamic desensitization strategies; the evaluation indicators include: data security, data availability, business compliance, and performance indicators; Establish user feedback channels to collect business users' experience and opinions on the use of desensitized data; Improve and optimize the reinforcement learning model based on evaluation metrics and user feedback; Regularly retrain the reinforcement learning model and incorporate new data and feedback information into the training process to continuously optimize the reinforcement learning model and dynamic desensitization strategy.

8. An artificial intelligence-based dynamic data adaptive desensitization device, implemented based on the artificial intelligence-based dynamic data adaptive desensitization method according to claim 1, characterized in that: include: Intelligent sensitive data identification module, used to mark sensitive information in dynamic data based on deep learning models and natural language processing technology; Adaptive desensitization algorithm module, used to train reinforcement learning models and generate dynamic desensitization strategies for dynamic data to perform data desensitization processing; The data association relationship maintenance module is used to identify and classify the association relationships in dynamic data through graph neural networks and perform unified processing on the associated data; The desensitized data quality optimization module is used to complete the desensitization processing of dynamic data, receive desensitization feedback data, and adjust the dynamic desensitization strategy based on the feedback desensitized data.

9. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the processor is connected to the memory via the bus, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, they are used to implement the dynamic data adaptive desensitization method based on artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a server, implements the artificial intelligence-based dynamic data adaptive desensitization method as described in any one of claims 1 to 7.

Images