Agent-based sensitive data grading method, device, equipment and storage medium

Abstract

The application discloses an agent-based sensitive data grading method and device, equipment and a storage medium, and relates to the technical field of information security. The method comprises the following steps: performing quality filtering on original data to be graded to obtain candidate sensitive data; pre-training and fine-tuning an original large language model through an industry corpus dataset to obtain a data security large model comprising a data classification agent and a data grading agent; calling the data classification agent to classify the candidate sensitive data based on a classification grading standard to obtain data category attributes; and calling the data grading agent to grade the candidate sensitive data according to the data category attributes and the classification grading standard to generate a classification grading list. The original large language model is pre-trained and fine-tuned through the industry corpus dataset, which provides the agents with the ability of natural language understanding and generation, and can adapt to sensitive data recognition in different scenarios, thereby improving the recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information security, and particularly relates to a sensitive data grading method and device based on an intelligent agent, equipment and a storage medium. BACKGROUND

[0002] With the rapid development of information technology, data is increasingly widely used in various fields. In enterprises and various organizations, the value of data is continuously increasing, but at the same time, the risk of sensitive data leakage is also increasing. Therefore, accurately defining, classifying and grading sensitive data has become an important task of data security management.

[0003] Current definitions, classifications and gradings of sensitive data generally have regular expression matching, keyword or dictionary matching, pattern matching and checking, and manual review marking. Regular expression matching is to identify sensitive data through a preset pattern, which is suitable for fixed format data. Keyword or dictionary matching is to search text based on a sensitive word library (such as "password", "confidential", "medical record", etc.) to identify data containing sensitive keywords. Pattern matching and checking is to combine format rules and checking algorithms to accurately identify sensitive data that meets specific checking logic. Manual review marking is to manually screen documents and database fields to directly mark sensitive data.

[0004] However, for sensitive data, the same data field may have different meanings and sensitivities in different scenarios. Therefore, the traditional fixed sensitive data identification method can generally only identify data with high standardization, resulting in low accuracy of sensitive data identification. SUMMARY

[0005] The main purpose of the present application is to provide a sensitive data grading method, device, equipment and storage medium based on an intelligent agent, which aims to solve the technical problem that the traditional fixed sensitive data identification method can generally only identify data with high standardization, resulting in low accuracy of sensitive data identification.

[0006] To achieve the above purpose, the present application provides a sensitive data grading method based on an intelligent agent, which comprises:

[0007] Filtering the data quality of the original data to be graded to obtain candidate sensitive data, wherein the original data to be graded is obtained from a multi-source database;

[0008] Pre-training and fine-tuning an original large language model through an industry corpus dataset to obtain a data security large model, wherein the data security large model comprises a data classification intelligent agent and a data grading intelligent agent;

[0009] based on the classification grading standard imported by the user, calling the data classification agent to classify the candidate sensitive data to obtain data category attributes;

[0010] According to the data category attribute and the classification grading standard, the data grading agent is called to grade the candidate sensitive data to generate a classification grading list.

[0011] In an embodiment, the data classification agent is composed of a Transformer encoder and a classifier, and the Transformer encoder is composed of a multi-head self-attention mechanism and a feedforward neural network. The step of calling the data classification agent to classify the candidate sensitive data based on the classification grading standard imported by the user to obtain data category attributes comprises:

[0012] Encode the importance of the candidate sensitive data through the multi-head self-attention mechanism to obtain the self-attention weight corresponding to each sequence data in the candidate sensitive data;

[0013] The self-attention weight is input into the feedforward neural network for linear transformation to obtain the data features corresponding to the candidate sensitive data;

[0014] Based on the data features and the classification grading standard imported by the user, the classifier is used to classify the candidate sensitive data to obtain data category attributes.

[0015] In an embodiment, the step of calling the data grading agent to grade the candidate sensitive data based on the data category attribute and the classification grading standard to generate a classification grading list comprises:

[0016] The data category attribute is used as a grading element of the data grading agent;

[0017] According to the data grading agent, the classification grading standard is parsed to determine the key elements and grading targets in the classification grading standard;

[0018] The grading element and the key element are associated and mapped to generate a classification grading mapping table;

[0019] Based on the grading target, the candidate sensitive data is graded according to the classification grading mapping table to generate a classification grading list.

[0020] In an embodiment, after the step of calling the data grading agent to grade the candidate sensitive data based on the data category attribute and the classification grading standard to generate a classification grading list, the method further comprises:

[0021] Receiving a user-inputted statement to be detected;

[0022] The data security big data model is used to perform data association on the statement to be detected, generating semantically related data.

[0023] Based on the semantic association data and the statement to be detected, a semantic search is performed in the classification and grading list to determine the sensitive data classification and grading of the statement to be detected.

[0024] In one embodiment, the step of pre-training and fine-tuning the original large language model using an industry corpus dataset to obtain a data security large model includes:

[0025] Collect industry-related linguistic datasets related to power grid knowledge;

[0026] The industry corpus dataset is transformed to obtain the corresponding natural language identification data.

[0027] Based on the natural language identifier data and the pre-set expert knowledge database, the original large language model is pre-trained and fine-tuned using LoRA technology to generate a data security large model. The original large language model is constructed using the Transformer architecture.

[0028] In one embodiment, the step of performing data quality filtering on the raw data to be graded to obtain candidate sensitive data includes:

[0029] Collect raw data to be classified from multiple source databases;

[0030] The raw data to be classified is filtered for keywords by a preset semantic recognition intelligent agent to obtain security topic corpus data;

[0031] The security-themed corpus data is cleaned to obtain cleaned corpus data;

[0032] The cleaned corpus data is filtered for quality using a preset fast text quality model and language model to obtain high-quality text data.

[0033] The high-quality text data is globally deduplicated to obtain candidate sensitive data corresponding to the original data to be classified.

[0034] In one embodiment, the step of performing quality filtering on the cleaned corpus data using a preset fast text quality model and language model to obtain high-quality text data includes:

[0035] The cleaned corpus data is input into a preset fast text quality model for feature matching to obtain initial high-quality data;

[0036] The initial high-quality data is input into a language model for probability prediction to determine the perplexity value of each text in the initial high-quality data;

[0037] Determine whether the confusion value is lower than a first preset threshold;

[0038] Text that falls below the first preset threshold is considered high-quality text data.

[0039] Furthermore, to achieve the above objectives, this application also proposes a sensitive data classification device based on intelligent agents, the device comprising:

[0040] The data filtering module is used to perform data quality filtering on the raw data to be classified to obtain candidate sensitive data, wherein the raw data to be classified is obtained from a multi-source database;

[0041] The model training module is used to pre-train and fine-tune the original large language model using an industry corpus dataset to obtain a large data security model, which includes a data classification agent and a data classification agent.

[0042] The data classification module is used to classify the candidate sensitive data based on the classification and grading standards imported by the user, and to obtain the data category attributes.

[0043] The data classification module is used to classify the candidate sensitive data according to the data category attributes and the classification and classification criteria, and generate a classification and classification list.

[0044] Furthermore, to achieve the above objectives, this application also proposes an agent-based sensitive data classification device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the agent-based sensitive data classification method described above.

[0045] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the agent-based sensitive data classification method described above.

[0046] One or more technical solutions proposed in this application have at least the following technical effects: The sensitive data classification method based on intelligent agents in this application includes: filtering the original data to be classified to obtain candidate sensitive data, wherein the original data to be classified is obtained from a multi-source database; pre-training and fine-tuning the original large language model using an industry corpus dataset to obtain a large data security model, wherein the large data security model includes a data classification intelligent agent and a data classification intelligent agent; based on the classification and grading standards imported by the user, calling the data classification intelligent agent to classify the candidate sensitive data to obtain data category attributes; and based on the data category attributes and the classification and grading standards, calling the data classification intelligent agent to classify the candidate sensitive data to generate a classification and grading list.

[0047] Because this application pre-trains and fine-tunes the original large language model based on industry corpus datasets, it provides natural language understanding and generation capabilities for data classification and data grading agents. This allows for accurate sensitive data classification and grading of candidate sensitive data according to classification and grading standards. This avoids the limitation of traditional sensitive data identification methods that can only identify standardized data, enabling it to adapt to sensitive data identification in different scenarios and thus improving the accuracy of identification. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating an embodiment of the sensitive data classification method based on intelligent agents in this application.

[0051] Figure 2 This is a flowchart illustrating Embodiment 2 of the sensitive data classification method based on intelligent agents in this application;

[0052] Figure 3 A schematic diagram of the data quality filtering process provided in Embodiment 2 of this application;

[0053] Figure 4 This is a flowchart illustrating Embodiment 3 of the sensitive data classification method based on intelligent agents in this application;

[0054] Figure 5A flowchart illustrating the overall process of identification and reasoning for the intelligent agent provided in Embodiment 3 of this application;

[0055] Figure 6 This is a schematic diagram of the module structure of the sensitive data classification device based on intelligent agents according to an embodiment of this application;

[0056] Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the sensitive data classification method based on intelligent agents in the embodiments of this application.

[0057] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0058] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0059] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0060] It should be noted that the executing entity in this embodiment can be a computing service device with data quality filtering, data classification, and data grading functions, such as a personal computer or a server, or an electronic device capable of performing the above functions, or an agent-based sensitive data grading device (hereinafter referred to as a data grading device) that executes the agent-based sensitive data grading method of this application. This embodiment does not limit this. The following uses a data grading device as an example to describe this embodiment and the following embodiments.

[0061] This application presents Embodiment 1. Embodiment 1 of this application provides a sensitive data classification method based on intelligent agents, referring to… Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the sensitive data classification method based on intelligent agents in this application.

[0062] In this embodiment, the sensitive data classification method based on intelligent agents includes steps S10 to S40:

[0063] Step S10: Perform data quality filtering on the raw data to be graded to obtain candidate sensitive data. The raw data to be graded is obtained from a multi-source database.

[0064] It should be noted that the raw data to be classified is an unprocessed initial dataset in its initial state. Due to the sensitive nature of data, the same data field may have different meanings and sensitivities in different scenarios. Therefore, this embodiment processes the raw data to be classified to adapt to the identification of sensitive data in different scenarios.

[0065] For example, consider current data in the power grid sector. In routine power grid dispatching, current data reflects the flow of electricity in transmission lines, serving as the basis for dispatchers to assess line load and ensure stable grid operation. Its sensitivity is moderate, and fluctuations within the power grid system are within normal monitoring limits. However, in power grid equipment upgrade planning, current data reflects the intensity of electricity demand in various regions, impacting crucial decisions such as equipment selection and layout. Its sensitivity is high, and data leakage or misinterpretation can affect the scientific and economic efficiency of power grid upgrades. Similarly, consider weight data in the medical field. In routine physical examinations, weight data reflects an individual's basic physical condition, used for preliminary health assessment. Its sensitivity is moderate, and privacy exists only between the examination institution and the individual. However, in drug clinical trials, weight data affects drug dosage calculations and grouping, making it a critical factor in ensuring the scientific accuracy of the trial. Its sensitivity is high, requiring strict confidentiality and precise management. The following explanation uses data from the power grid sector; this embodiment does not impose limitations.

[0066] Understandably, multi-source databases originate from different systems, platforms, or data repositories. This data is stored in various systems within the power grid system in different formats and structures, such as relational databases, time-series databases, and business systems.

[0067] It should be understood that candidate sensitive data is data selected from the original data to be classified after data quality filtering (such as data cleaning, deduplication, etc.).

[0068] In practice, the data classification device collects data from multiple source databases within the power grid system, such as relational databases, time-series databases, and business systems, to obtain the raw data to be classified. This raw data is then cleaned and deduplicated to obtain candidate sensitive data.

[0069] Step S20: Pre-train and fine-tune the original large language model using an industry corpus dataset to obtain a large data security model, which includes a data classification agent and a data classification agent.

[0070] It should be noted that the industry corpus dataset is a collection of text data related to the power grid sector. This dataset contains various terms, typical expressions, and business logic related to the power grid. Using the industry corpus dataset for pre-training and fine-tuning when building a large-scale data security model allows the model to better understand the characteristics and needs of the power grid sector, thereby improving the model's accuracy in data classification and grading for power grid data.

[0071] Understandably, the data security big model is an artificial intelligence model specifically built for the field of data security. It includes two important parts: a data classification agent and a data classification agent. Its main function is to classify and classify the aforementioned candidate sensitive data.

[0072] The data classification agent's function is to classify candidate sensitive data, identifying different characteristics of the candidate sensitive data and categorizing them into different data category attributes. The data rating agent can accurately assign appropriate levels to candidate sensitive data based on factors such as data category attributes, data sensitivity, and importance.

[0073] In the specific implementation, after obtaining the above-mentioned candidate sensitive data, it is also necessary to pre-train and fine-tune the original large language model using industry corpus datasets related to the power grid field, so as to obtain a data classification agent that can identify different features of the candidate sensitive data and classify them into different data category attributes, and a data rating agent that can assign corresponding levels to the candidate sensitive data according to data category attributes, data sensitivity and other factors.

[0074] In another feasible implementation, step S20 of this embodiment may include the following steps: collecting an industry corpus dataset related to power grid knowledge; performing data transformation on the industry corpus dataset to obtain corresponding natural language identification data; and, based on the natural language identification data and a preset expert knowledge database, pre-training and fine-tuning the original large language model using LoRA technology to generate a large data security model, wherein the original large language model is constructed using the Transformer architecture.

[0075] It should be noted that natural language labeling data is a data format that identifies information in the power grid industry corpus in a natural language manner. For example, the parameters of power grid equipment might originally be a string of numbers and codes, but after conversion, they might become a description in natural language form such as "a certain type of transformer, with a rated power of X kilowatts and a voltage level of Y kilovolts". This data format can be understood and processed by large language models, providing a foundation for subsequent model training using this data.

[0076] Understandably, an expert knowledge database is a database that compiles the knowledge of experts in the power grid field. The content may include experts' insights into complex power grid problems, unique fault diagnosis methods, and predictions of power grid development trends.

[0077] It should be understood that LoRA (Low-Rank Adaptation) is a model optimization technique primarily used for pre-training and fine-tuning of original large language models built on the Transformer architecture. LoRA introduces low-rank matrix factorization into the attention mechanism layer of the original large language model, enabling efficient updating and optimization of model parameters. This allows the model to better adapt to the specific semantic and data characteristics of the power grid industry without significantly increasing computational resources and training time. Specifically, LoRA freezes most of the parameters of the original large language model, training only on the decomposed low-rank matrix, reducing the number of parameters and computational cost while preserving the general language knowledge and semantic understanding capabilities of the original large language model. During fine-tuning, overfitting is prevented and the model's generalization performance is improved by adjusting hyperparameters such as the learning rate and batch size, and employing strategies such as early stopping and regularization. Furthermore, to further optimize the model, techniques such as knowledge distillation can be used to transfer the knowledge and performance of the original large language model to a smaller model, reducing deployment costs and inference time, making it more suitable for operation in real-world industry applications.

[0078] Among them, the Transformer architecture has advantages such as self-attention mechanism, which can effectively process long sequence data. This makes the original large language model perform well in natural language processing and has a certain language understanding and generation capabilities.

[0079] In this implementation, a large-scale data security model is obtained by collecting a corpus dataset related to power grid knowledge and converting it into natural language tagging data. This data is then combined with an expert knowledge database and LoRA technology is used to pre-train and fine-tune the original large language model based on the Transformer architecture. This model is highly industry-specific and can accurately understand security matters related to the characteristics of power grid data.

[0080] Step S30: Based on the classification and grading standards imported by the user, call the data classification agent to classify the candidate sensitive data and obtain the data category attributes.

[0081] Understandably, classification and grading standards are a set of standards imported by users that clearly define how to classify and grade data. These standards serve as the basis for data classification and grading agents to perform classification and grading. For example, different category and level division rules can be formulated based on factors such as the content type of data in the power grid sector, the entities involved, and the degree of impact on business.

[0082] It should be noted that data category attributes are identifiers assigned after classifying data based on its essential characteristics, reflecting the type of candidate sensitive data within the classification and grading standard system.

[0083] In practical implementation, after completing the pre-training and fine-tuning of the above model, the classification and grading standards and candidate sensitive data can be imported into the data classification agent to classify the data and obtain the data category attributes.

[0084] In one feasible implementation, the data classification agent consists of a Transformer encoder and a classifier. The Transformer encoder consists of a multi-head self-attention mechanism and a feedforward neural network. Step S30 in this embodiment may include the following steps: encoding the importance of the candidate sensitive data through the multi-head self-attention mechanism to obtain the self-attention weights corresponding to each sequence of data in the candidate sensitive data; inputting the self-attention weights into the feedforward neural network for linear transformation to obtain the data features corresponding to the candidate sensitive data; and classifying the candidate sensitive data for sensitive words through the classifier based on the data features and the classification and grading standards imported by the user to obtain the data category attributes.

[0085] It should be noted that the data classification agent consists of a Transformer encoder and a classifier. The Transformer encoder comprises multiple encoding layers, each consisting of a multi-head self-attention mechanism and a feedforward neural network. The main function of the Transformer encoder is to encode candidate sensitive data. Through the multi-head self-attention mechanism, it can simultaneously determine the correlation between each position in the candidate sensitive data and other positions from multiple different representation subspaces, i.e., the self-attention weight (used to measure the relative importance between different sequences of data in the candidate sensitive data). Then, the feedforward neural network receives the self-attention weights output by the multi-head self-attention mechanism as input, performs linear transformations and other operations on the input, and calculates the input data through the connection weights between neurons, thereby further transforming the self-attention-weighted candidate sensitive data into data features more suitable for classification operations.

[0086] The Transformer's self-attention mechanism allows the model to dynamically focus on different positions in a sequence when processing sequential data, capturing long-distance dependencies in the data. This mechanism enables the Transformer to perform exceptionally well when processing natural language data, understanding the relationships and semantics between different words in a sentence.

[0087] A classifier is a component in a data classification agent used to perform the final classification of data. Examples include linear classifiers (such as logistic regression) and nonlinear classifiers (such as multilayer perceptrons and support vector machines). Based on the data features output by the Transformer encoder and the classification and grading criteria imported by the user, the classifier performs classification operations on candidate sensitive data, dividing the data into different categories and thus determining the data category attributes.

[0088] Traditional data classification techniques, such as purely manual annotation and rule-based or dictionary-based data recognition, have many limitations. However, the data classification technology based on the intelligent agent in this embodiment offers the following advantages: leveraging the strong generalization ability of a large language model, it can cover all data features after learning from a small amount of sample data. In practical applications, the intelligent agent data classification technology based on a large language model has demonstrated enormous potential and value. Taking industries such as meteorology, energy, and healthcare as examples, this technology can be used for the automatic classification and management of structured and unstructured data. By understanding the natural language descriptions in business processes, the model can classify application data into basic, sensitive, and confidential categories, or it can categorize data according to major industry labels, greatly improving the efficiency and accuracy of industry data management. In the financial sector, it can be used for the classification and risk assessment of transaction data. The model can analyze transaction descriptions, amounts, times, and other information to identify abnormal transaction patterns, helping financial institutions prevent fraud and compliance risks.

[0089] In this implementation, the Transformer encoder (containing a multi-head self-attention mechanism and a feedforward neural network) in the data classification agent can comprehensively and deeply extract the features of candidate sensitive data. These features are then classified by a classifier based on user-imported classification and grading standards, achieving effective data feature mining. This not only improves classification accuracy but also provides flexible classification criteria, adapting to various scenarios and enabling effective management of sensitive data.

[0090] Step S40: Based on the data category attributes and the classification and grading standards, call the data grading agent to grade the candidate sensitive data and generate a classification and grading list.

[0091] It should be noted that the classification and grading list is a document generated after the data has been classified and graded. It details the name or identifier of the data, the corresponding category attribute, and the determined level.

[0092] In the specific implementation, after outputting the above data category attributes, the data category attributes and classification and grading standards are then input into the data grading agent to classify the candidate sensitive data and generate a classification and grading list.

[0093] In another feasible implementation, step S40 of this embodiment may include the following steps: using the data category attribute as a classification element of the data classification agent; parsing the classification and grading standards according to the data classification agent to determine the key elements and grading targets in the classification and grading standards; associating and mapping the classification elements and the key elements to generate a classification and grading mapping table; and classifying the candidate sensitive data according to the classification and grading mapping table based on the grading targets to generate a classification and grading list.

[0094] It's important to note that key elements are the decisive components of the classification and grading standards. For example, for personal information data, this might involve identifying the scope of personnel involved (e.g., all employees, senior management, etc.) and the data update frequency. The grading objective is the goal of classifying candidate sensitive data, which is to apply different levels of security protection measures to different levels of data, such as the data's sensitivity level (e.g., top secret, confidential, secret, etc.) and the core nature of the business (e.g., core business, important business, general business, etc.).

[0095] Understandably, the classification and grading mapping table is a table that corresponds data category attributes to key elements in the classification and grading criteria. For example, if the data category is "core business data" and it has high-value, high-risk attributes (according to the previously defined grading elements), and the user template specifies that such high-value, high-risk data is at level L3, then a mapping relationship of "core business data - L3" is established. At the same time, considering the correlation between data, some data, although appearing to belong to a low level individually, may need to have their level adjusted if they are related to other high-level data (such as serving as auxiliary data for high-level data).

[0096] In this implementation, through clear data category attribute definitions and the operation of the data classification agent, candidate sensitive data can be accurately classified and classified according to classification and grading standards. Secondly, by parsing the classification and grading standards to determine key elements and grading objectives, the classification and grading process becomes more targeted and better meets the needs of data security management and compliance. Furthermore, the generated classification and grading mapping table provides a clear basis for data grading, improving the efficiency of the grading process.

[0097] In the technical solution provided in this embodiment, the data classification device collects data from multiple source databases of the power grid system, such as relational databases, time-series databases, and business systems, to obtain the raw data to be classified. Then, the raw data to be classified undergoes data cleaning and deduplication processes to obtain candidate sensitive data. Next, the original large language model is pre-trained and fine-tuned using industry-related corpus datasets in the power grid field, resulting in a data classification agent capable of identifying different features of the candidate sensitive data and classifying them into different data category attributes, and a data classification agent capable of assigning appropriate levels to the candidate sensitive data based on data category attributes, data sensitivity, and other factors. After completing the pre-training and fine-tuning of the above models, the classification and grading standards and candidate sensitive data can be imported into the data classification agent for data classification to obtain data category attributes. Finally, the data category attributes and classification and grading standards are input into the data classification agent to classify the candidate sensitive data and generate a classification and grading list. Because this embodiment pre-trains and fine-tunes the original large language model based on an industry corpus dataset, it provides natural language understanding and generation capabilities for the data classification agent and the data grading agent. This allows for accurate sensitive data classification and grading of candidate sensitive data according to classification and grading standards. This avoids the limitation of traditional sensitive data identification methods that can only identify standardized data, and enables the identification of sensitive data in different scenarios, thereby improving the accuracy of identification.

[0098] Based on the first embodiment of this application described above, a second embodiment of this application is proposed. In this second embodiment, content that is the same as or similar to that in the first embodiment can be referred to the above description and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating Embodiment 2 of the sensitive data classification method based on intelligent agents in this application.

[0099] Step S10 in this example includes steps S11 to S15:

[0100] Step S11: Collect the raw data to be classified from the multi-source database.

[0101] Step S12: Use a preset semantic recognition agent to filter keywords in the raw data to be classified to obtain security topic corpus data.

[0102] It should be noted that the semantic recognition agent is an intelligent agent that can understand the semantics of text. It can analyze the input raw data to be classified, identify the keywords in the text of the raw data to be classified, and accurately filter out keywords related to the security topic by analyzing the semantic relationship of words, context and other factors. Thus, the data containing these keywords is filtered out to obtain the security topic corpus data.

[0103] Specifically, keyword filtering is divided into two parts: secure keyword retrieval and retrieval by a semantic recognition agent. Secure keyword retrieval: This involves matching and filtering the raw data to be classified using predefined related keywords to quickly locate data fragments containing these keywords. However, considering that keyword retrieval alone may not be able to capture contextual semantics and semantic relationships, subsequent retrieval by a semantic recognition agent is also necessary.

[0104] In the retrieval of semantic recognition agents, pre-trained semantic recognition agents are used to perform semantic understanding and relevance ranking of data, and to mine corpora that are highly relevant to the security topic, even if they do not fully contain the preset keywords.

[0105] The raw data to be classified is converted into semantic vector representations. Techniques such as the multi-head attention mechanism in the Transformer architecture are used to capture long-distance dependencies for better semantic understanding. Then, a relevance score is obtained by calculating the distance (e.g., cosine distance) between the semantic vectors and security topic vectors (i.e., the predefined related keywords mentioned above). The semantic vectors are then ranked according to the scores, thereby achieving semantic understanding and relevance ranking of the raw data to be classified, and mining security topic corpus data highly relevant to security topics.

[0106] Step S13: Perform data cleaning on the security-themed corpus data to obtain cleaned corpus data.

[0107] Specifically, the security-themed corpus data retrieved above often suffers from noise, duplication, and inconsistency. Therefore, careful data cleaning is necessary before it can be used for subsequent model training. Data cleaning mainly includes the following aspects: First, local deduplication is performed by comparing unique identifiers and content fingerprints to identify and remove duplicate records, avoiding redundancy and misleading effects on model training. Second, data that does not conform to the target language is filtered out to ensure linguistic consistency of the corpus. Third, operations such as correcting spelling errors, standardizing formatting, and filling in missing values ​​are performed to obtain cleaned corpus data, further improving data quality and consistency, and providing a clean and reliable corpus foundation for model training.

[0108] Step S14: The cleaned corpus data is filtered for quality using a preset fast text quality model and language model to obtain high-quality text data.

[0109] It's important to note that the Fast Text Quality Model (FTM) is a model used to evaluate text quality. Trained based on the Fast Text algorithm, it can assess and classify the quality of text data. By learning the features of a large amount of high-quality and low-quality text, the FTM establishes quality assessment criteria, thereby automatically selecting corpora that meet the quality requirements.

[0110] Understandably, the language model follows the same principles as the original Large Language Model (LLM) described above. It can be pre-trained on large amounts of text data and is capable of handling a wide variety of natural language processing tasks. The language model can use either a large or small model; this embodiment does not impose any restrictions. The language model measures the data from the perspectives of fluency and semantic coherence, calculating the perplexity score of the text. A lower perplexity score indicates that the text better matches the language model's predictions and has more logical and coherent meaning, thus yielding high-quality text data.

[0111] In one feasible implementation, step S14 of this example includes the following steps: inputting the cleaned corpus data into a preset fast text quality model for feature matching to obtain initial high-quality data; inputting the initial high-quality data into a language model for probability prediction to determine the perplexity value of each text in the initial high-quality data; determining whether the perplexity value is lower than a first preset threshold; and taking the text corresponding to the value lower than the first preset threshold as high-quality text data.

[0112] It's important to note that the initial high-quality data is text data initially identified as having relatively high quality after feature matching by a fast text quality model. For the cleaned corpus data, the fast text quality model extracts its features and matches them with the feature patterns of high-quality and low-quality text obtained during training to determine the initial high-quality data within the cleaned corpus. For example, if the cleaned corpus data contains a large number of rare, potentially misspelled words, and the sentence structure does not conform to common grammatical patterns, then it is likely closer to the feature patterns of low-quality text.

[0113] Understandably, perplexity is a measure of the uncertainty a language model faces when predicting text, reflecting the difficulty the language model faces in predicting each text in the initial high-quality data. A lower perplexity value indicates that the language model's prediction of the text is relatively accurate, suggesting that the text's structure, vocabulary usage, etc., conform to the rules learned by the language model, which means that the text has high quality; while a higher perplexity value indicates that the text may have some problems such as not conforming to language habits or semantic incoherence.

[0114] The first preset threshold is a pre-defined numerical limit used to distinguish between high and low text quality. When the perplexity value of text is lower than the first preset threshold, the text is considered to be of high quality and can be accepted as high-quality text data; conversely, when the perplexity value is higher than the first preset threshold, the text is considered to have potential problems and does not meet the requirements of high-quality text. The setting of the first preset threshold can be determined based on specific task requirements, corpus characteristics, and experience; this embodiment does not impose any restrictions on it.

[0115] In this implementation, the initial screening using a fast text quality model can quickly remove text of obviously poor quality, improving processing efficiency. Then, by utilizing a language model for more in-depth probability prediction and perplexity calculation, text quality can be evaluated from a more comprehensive perspective, including semantics and language habits, making the selected high-quality text data more accurate and reliable.

[0116] Step S15: Perform global deduplication on the high-quality text data to obtain candidate sensitive data corresponding to the original data to be classified.

[0117] It should be noted that, to further ensure data uniqueness and conciseness, global deduplication can be performed. First, by extracting the minimum hash value, the similarity between high-quality text data can be quickly estimated, efficiently identifying similar records in large-scale datasets. Then, the LSH (Locality-Sensitive Hashing) algorithm is used to map high-dimensional data to a low-dimensional space, making similar data more likely to be grouped into the same bucket in the low-dimensional space. This effectively removes duplicate data globally, providing a concise and efficient dataset for model training.

[0118] For example, to facilitate understanding of the above-mentioned processing procedure for the raw data to be classified, refer to... Figure 3 , Figure 3 This is a flowchart illustrating the data quality filtering process provided in Embodiment 2 of this application. First, large-scale data collection is performed to obtain the raw data to be classified. Then, keyword filtering is performed on the collected raw data to obtain security-themed corpus data. Next, the security-themed corpus data is cleaned to obtain cleaned corpus data. Then, data quality filtering and global deduplication are performed on the cleaned corpus data to obtain candidate sensitive data corresponding to the raw data to be classified. Subsequently, data is extracted from the candidate sensitive data for evaluation to determine whether it conforms to the security theme, thus verifying the validity and reliability of the corpus.

[0119] In the technical solution provided in this embodiment, data processing is carried out in multiple stages, including data collection, security topic corpus retrieval, data cleaning, data quality filtering, and global deduplication. The final candidate sensitive data has undergone a strict screening and processing process, which is more in line with the requirements of security management.

[0120] Based on the above embodiments of this application, a third embodiment of this application is proposed. In this third embodiment, content that is the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 , Figure 4 This is a flowchart illustrating Embodiment 3 of the sensitive data classification method based on intelligent agents in this application.

[0121] This example includes steps S51 to S53 after step S40:

[0122] Step S51: Receive the statement to be detected input by the user.

[0123] It should be noted that the statement to be detected refers to the statement entered by the user that needs to be checked for data security. It can be any text content, such as comments entered by the user on the power grid system, messages posted, or document content submitted by the user.

[0124] Step S52: Perform data association on the statement to be detected using the data security big data model to generate semantic association data.

[0125] It should be noted that semantic association data is the result data obtained through data association operations. It contains various information that has semantic association with the statement to be detected, such as other words, phrases, sentences, or complete document content.

[0126] When using the data security big data model for association, the model generates relevant specific data instances based on its pre-trained knowledge and semantic understanding capabilities, according to the category or level of the input. For example, if the input is "personal identification information," the data security big data model might suggest specific content such as "ID number," "name," and "home address."

[0127] Step S53: Based on the semantic association data and the statement to be detected, perform semantic retrieval in the classification and grading list to determine the sensitive data classification and grading of the statement to be detected.

[0128] It should be noted that the sensitive data classification and grading criteria are determined by finding the sensitive data classification and grading that best matches the statement to be detected from the classification and grading list. This method allows for a more accurate determination of whether the statement to be detected contains sensitive data and the importance of that sensitive data.

[0129] In the technical solution provided in this embodiment, after receiving the user's input statement to be detected, the statement is first identified. If it is a specific data description (such as a mobile phone number), it is directly retrieved; if it is a data category (such as personal identification information) or data level, a large data security model is used to perform data association and generate semantically related data. Finally, semantically related data is used to perform semantic retrieval in a classification and grading list. Once the corresponding sensitive data is retrieved, the sensitive data classification and grading can be located.

[0130] For example, to help understand the implementation process of the agent-based sensitive data classification method obtained by combining Embodiment 1 and Embodiment 3 above, please refer to... Figure 5 ,Figure 5 This is an overall flowchart of the identification and reasoning process of the intelligent agent provided in Embodiment 3 of this application; specifically:

[0131] First, the classification and grading standards are input into the "Classification Standard Understanding Agent" and the "Grading Standard Understanding Agent" respectively to understand the standards. Then, the understanding results are input into the "Data Classification Agent" and the "Data Grading Agent" respectively.

[0132] Then, after obtaining the raw data to be classified, the raw data to be classified is input into the "semantic recognition agent" to understand its meaning, the "data classification agent" to classify it according to the established standards, and the "data classification agent" to determine the sensitivity level, thereby completing the three steps of "data semantic recognition route", "data classification route" and "data level recognition route", and finally outputting the classification and classification results.

[0133] Subsequently, after the user inputs the data to be detected, the "semantic retrieval agent" efficiently retrieves and locates the data to be detected, forms a "data location route", and outputs the distribution of sensitive data, thereby clarifying the distribution of sensitive data in the entire data to be detected, and providing a basis for subsequent security management.

[0134] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the sensitive data classification method based on intelligent agents in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0135] This application also provides a sensitive data classification device based on intelligent agents; please refer to... Figure 6 , Figure 6 This is a schematic diagram of the module structure of a sensitive data classification device based on intelligent agents according to an embodiment of this application; the sensitive data classification device based on intelligent agents includes:

[0136] Data filtering module 601 is used to perform data quality filtering on the raw data to be classified to obtain candidate sensitive data, wherein the raw data to be classified is obtained from a multi-source database;

[0137] The model training module 602 is used to pre-train and fine-tune the original large language model using an industry corpus dataset to obtain a large data security model, which includes a data classification agent and a data classification agent.

[0138] The data classification module 603 is used to call the data classification agent to classify the candidate sensitive data based on the classification and grading standards imported by the user, and obtain the data category attributes.

[0139] The data classification module 604 is used to call the data classification agent to classify the candidate sensitive data according to the data category attributes and the classification and classification standards, and generate a classification and classification list.

[0140] In one implementation, the data classification agent consists of a Transformer encoder and a classifier. The Transformer encoder consists of a multi-head self-attention mechanism and a feedforward neural network. The data classification module 603 is further configured to encode the importance of the candidate sensitive data through the multi-head self-attention mechanism to obtain the self-attention weights corresponding to each sequence of data in the candidate sensitive data; input the self-attention weights into the feedforward neural network for linear transformation to obtain the data features corresponding to the candidate sensitive data; and based on the data features and the classification and grading standards imported by the user, classify the candidate sensitive data for sensitive words through the classifier to obtain the data category attributes.

[0141] In one implementation, the data classification module 604 is further configured to use the data category attribute as a classification element of the data classification agent; based on the data classification agent, parse the classification and grading standards to determine the key elements and grading targets in the classification and grading standards; associate and map the classification elements and the key elements to generate a classification and grading mapping table; and based on the grading targets, classify the candidate sensitive data according to the classification and grading mapping table to generate a classification and grading list.

[0142] As one implementation, the agent-based sensitive data classification device is further configured to receive a user-inputted statement to be detected; perform data association on the statement to be detected through the data security big data model to generate semantic association data; and perform semantic retrieval in the classification and classification list based on the semantic association data and the statement to be detected to determine the sensitive data classification and classification of the statement to be detected.

[0143] As one implementation, the model training module 602 is also used to collect industry corpus datasets related to power grid knowledge; perform data transformation on the industry corpus datasets to obtain corresponding natural language identification data; and, based on the natural language identification data and a preset expert knowledge database, pre-train and fine-tune the original large language model using LoRA technology to generate a large data security model, wherein the original large language model is constructed using the Transformer architecture.

[0144] In one implementation, the data filtering module 601 is further configured to: collect raw data to be classified from a multi-source database; filter keywords in the raw data to be classified using a preset semantic recognition agent to obtain security-themed corpus data; clean the security-themed corpus data to obtain cleaned corpus data; perform quality filtering on the cleaned corpus data using a preset fast text quality model and language model to obtain high-quality text data; and perform global deduplication on the high-quality text data to obtain candidate sensitive data corresponding to the raw data to be classified.

[0145] In one implementation, the data filtering module 601 is further configured to input the cleaned corpus data into a preset fast text quality model for feature matching to obtain initial high-quality data; input the initial high-quality data into a language model for probability prediction to determine the perplexity value of each text in the initial high-quality data; determine whether the perplexity value is lower than a first preset threshold; and take the text corresponding to the value lower than the first preset threshold as high-quality text data.

[0146] Other embodiments or specific implementations of the sensitive data classification device based on intelligent agents in this application can be found in the above-described method embodiments, and will not be repeated here.

[0147] The sensitive data classification device based on intelligent agents provided in this application, employing the sensitive data classification method based on intelligent agents in the above embodiments, can solve the technical problem that traditional fixed sensitive data identification methods can generally only identify data with a high degree of standardization, resulting in low accuracy in sensitive data identification. Compared with the prior art, the beneficial effects of the sensitive data classification device based on intelligent agents provided in this application are the same as those of the sensitive data classification method based on intelligent agents provided in the above embodiments, and other technical features in the sensitive data classification device based on intelligent agents are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0148] This application provides a sensitive data classification device based on intelligent agents. The sensitive data classification device based on intelligent agents includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the sensitive data classification method based on intelligent agents in the first embodiment described above.

[0149] The following is for reference. Figure 7 , Figure 7This is a schematic diagram of the device structure of the hardware operating environment involved in the sensitive data classification method based on intelligent agents in the embodiments of this application. It shows a schematic diagram of the structure of the device suitable for implementing the sensitive data classification method based on intelligent agents in the embodiments of this application. The sensitive data classification device based on intelligent agents in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), vehicle terminals (such as vehicle navigation terminals), etc., as well as fixed terminals such as digital TVs, desktop computers, etc. Figure 7 The agent-based sensitive data classification device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0150] like Figure 7 As shown, the agent-based sensitive data classification device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the agent-based sensitive data classification device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the agent-based sensitive data classification device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows agent-based sensitive data classification devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0151] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0152] The sensitive data classification device based on intelligent agents provided in this application, employing the sensitive data classification method based on intelligent agents in the above embodiments, can solve the technical problem that traditional fixed sensitive data identification methods can generally only identify data with a high degree of standardization, resulting in low accuracy in sensitive data identification. Compared with the prior art, the beneficial effects of the sensitive data classification device based on intelligent agents provided in this application are the same as those of the sensitive data classification method based on intelligent agents provided in the above embodiments, and other technical features in this sensitive data classification device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0153] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0154] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0155] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the agent-based sensitive data classification method in the above embodiments.

[0156] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0157] The aforementioned computer-readable storage medium may be included in an agent-based sensitive data classification device; or it may exist independently and not assembled into an agent-based sensitive data classification device.

[0158] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by an agent-based sensitive data classification device, the agent-based sensitive data classification device performs the following actions: performs data quality filtering on the raw data to be classified to obtain candidate sensitive data, wherein the raw data to be classified is obtained from a multi-source database; pre-trains and fine-tunes the original large language model using an industry corpus dataset to obtain a large data security model, wherein the large data security model includes a data classification agent and a data classification agent; based on user-imported classification and classification standards, calls the data classification agent to classify the candidate sensitive data to obtain data category attributes; and, based on the data category attributes and the classification and classification standards, calls the data classification agent to classify the candidate sensitive data, generating a classification and classification list.

[0159] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0160] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0161] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0162] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described agent-based sensitive data classification method. This solves the technical problem that traditional, fixed sensitive data identification methods generally only identify data with a high degree of standardization, resulting in low accuracy in sensitive data identification. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the agent-based sensitive data classification method provided in the above embodiments, and will not be repeated here.

[0163] The above description is only a part of the embodiments of this application and does not limit the scope of protection of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the scope of protection of this application.

Claims

1. A sensitive data classification method based on intelligent agents, characterized in that, The method includes: Data quality filtering is performed on the raw data to be graded to obtain candidate sensitive data, wherein the raw data to be graded is obtained from a multi-source database; The original large language model is pre-trained and fine-tuned using an industry corpus dataset to obtain a large data security model, which includes a data classification agent and a data classification agent. Based on the classification and grading standards imported by the user, the data classification agent is invoked to classify the candidate sensitive data and obtain the data category attributes; Based on the data category attributes and the classification and grading standards, the data grading agent is invoked to grade the candidate sensitive data and generate a classification and grading list; The data classification agent consists of a Transformer encoder and a classifier. The Transformer encoder is composed of a multi-head self-attention mechanism and a feedforward neural network. The step of calling the data classification agent to classify the candidate sensitive data based on the user-imported classification and grading criteria to obtain the data category attributes includes: The importance of the candidate sensitive data is encoded by the multi-head self-attention mechanism to obtain the self-attention weights corresponding to each sequence of data in the candidate sensitive data. The self-attention weights are input into the feedforward neural network and linearly transformed to obtain the data features corresponding to the candidate sensitive data. Based on the data features and the user-imported classification and grading standards, the classifier is used to classify the candidate sensitive data for sensitive words to obtain data category attributes; The step of calling the data grading agent to grade the candidate sensitive data and generate a grading list based on the data category attributes and the classification and grading criteria includes: The data category attribute is used as the classification element of the data classification agent; Based on the data classification agent, the classification and grading criteria are analyzed to determine the key elements and grading objectives in the classification and grading criteria; The classification elements and key elements are associated and mapped to generate a classification and grading mapping table. Based on the grading objective, the candidate sensitive data are graded according to the classification and grading mapping table to generate a classification and grading list.

2. The method as described in claim 1, characterized in that, After the step of calling the data classification agent to classify the candidate sensitive data and generate a classification list based on the data category attributes and the classification and grading criteria, the method further includes: Receive the user-input statement to be tested; The data security big data model is used to perform data association on the statement to be detected, generating semantically related data. Based on the semantic association data and the statement to be detected, a semantic search is performed in the classification and grading list to determine the sensitive data classification and grading of the statement to be detected.

3. The method as described in claim 1, characterized in that, The steps of pre-training and fine-tuning the original large language model using an industry corpus dataset to obtain a data security large model include: Collect industry-related linguistic datasets related to power grid knowledge; The industry corpus dataset is transformed to obtain the corresponding natural language identification data. Based on the natural language identifier data and the pre-set expert knowledge database, the original large language model is pre-trained and fine-tuned using LoRA technology to generate a data security large model. The original large language model is constructed using the Transformer architecture.

4. The method according to any one of claims 1 to 3, characterized in that, The step of performing data quality filtering on the raw data to be graded to obtain candidate sensitive data includes: Collect raw data to be classified from multiple source databases; The raw data to be classified is filtered for keywords by a preset semantic recognition intelligent agent to obtain security topic corpus data; The security-themed corpus data is cleaned to obtain cleaned corpus data; The cleaned corpus data is filtered for quality using a preset fast text quality model and language model to obtain high-quality text data. The high-quality text data is globally deduplicated to obtain candidate sensitive data corresponding to the original data to be classified.

5. The method as described in claim 4, characterized in that, The step of performing quality filtering on the cleaned corpus data using a preset fast text quality model and language model to obtain high-quality text data includes: The cleaned corpus data is input into a preset fast text quality model for feature matching to obtain initial high-quality data; The initial high-quality data is input into a language model for probability prediction to determine the perplexity value of each text in the initial high-quality data; Determine whether the confusion value is lower than a first preset threshold; Text that falls below the first preset threshold is considered high-quality text data.

6. A sensitive data classification device based on intelligent agents, characterized in that, The device includes: The data filtering module is used to perform data quality filtering on the raw data to be classified to obtain candidate sensitive data, wherein the raw data to be classified is obtained from a multi-source database; The model training module is used to pre-train and fine-tune the original large language model using an industry corpus dataset to obtain a large data security model, which includes a data classification agent and a data classification agent. The data classification module is used to classify the candidate sensitive data based on the classification and grading standards imported by the user, and to obtain the data category attributes. The data classification module is used to call the data classification agent to classify the candidate sensitive data according to the data category attributes and the classification and classification criteria, and generate a classification and classification list. The data classification agent consists of a Transformer encoder and a classifier, and the Transformer encoder consists of a multi-head self-attention mechanism and a feedforward neural network. The data classification module is further configured to encode the importance of the candidate sensitive data through the multi-head self-attention mechanism to obtain the self-attention weights corresponding to each sequence of data in the candidate sensitive data; input the self-attention weights into the feedforward neural network for linear transformation to obtain the data features corresponding to the candidate sensitive data; and classify the candidate sensitive data for sensitive words through the classifier based on the data features and the classification and grading standards imported by the user to obtain the data category attributes. The data classification module is further configured to use the data category attributes as classification elements of the data classification agent; based on the data classification agent, analyze the classification and grading standards to determine the key elements and classification objectives in the classification and grading standards; associate and map the classification elements and the key elements to generate a classification and grading mapping table; and based on the classification objectives, classify the candidate sensitive data according to the classification and grading mapping table to generate a classification and grading list.

7. A sensitive data classification device based on intelligent agents, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the agent-based sensitive data classification method as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the agent-based sensitive data classification method as described in any one of claims 1 to 5.

Images