A method for classifying and grading sensitive personal information using knowledge graphs

Abstract

This invention relates to the field of information technology and provides a method for classifying and grading sensitive personal information using a knowledge graph. The aim is to address the problem that existing methods for classifying and grading sensitive personal information often only consider a single scenario and neglect the implementation of classification and grading across multiple scenarios. The main scheme includes: extracting entities and relationships from text data containing personal information to obtain sets of entities and relationships; constructing a personal information knowledge graph based on these sets; extracting features from each entity to obtain a corresponding feature vector; clustering the feature vectors to obtain clustering results; classifying each clustering category by security level to obtain sensitivity level information; and associating the sensitivity level information with the personal information knowledge graph to obtain a classification and grading knowledge graph for sensitive personal information. When new scenarios arise, incremental updates are performed using multi-knowledge graph fusion technology.

Description

Technical Field

[0001] This invention relates to the field of information technology and provides a method for classifying and grading sensitive personal information using knowledge graphs. Background Technology

[0002] In recent years, with the continuous development of technologies such as the Internet, big data, and artificial intelligence, while people enjoy the benefits brought by massive amounts of information, they also suffer from the problems caused by the leakage of sensitive personal information. To protect the security of sensitive personal information, relevant laws have been promulgated and implemented to strengthen the protection of personal data resources, establish a big data security management system, implement classified and graded management of data resources, and ensure safe, efficient, and reliable applications. For example, the Personal Information Protection Law has established strict standards for the use of personal information, making clear requirements for the collection, use, and storage of sensitive personal data, and stipulating the establishment of a compliance system for the protection of sensitive personal data, the establishment of specific institutions for supervision, regulation, and governance, and ensuring open, fair, and just rules. Article 21 of the Data Security Law of the People's Republic of China clearly states that the state establishes a classified and graded protection system for data, classifying and grading data according to its importance in economic and social development, and the degree of harm it would cause to national security, public interests, or the legitimate rights and interests of individuals and organizations if it is tampered with, damaged, leaked, or illegally obtained or used.

[0003] However, despite the promulgation of relevant laws by the state, the specific implementation of the classification and grading method still encounters several problems.

[0004] (1) Nowadays, various online applications, such as social media platforms, e-commerce platforms, and ride-hailing platforms, are filled with a large amount of sensitive personal information. The scenarios involved in sensitive personal information are wide-ranging, and how to classify and grade it in multiple scenarios is an unsolved problem. The 14 national cybersecurity standards promulgated on October 14, 2022, only protect single scenarios, such as facial information scenarios, voiceprint information scenarios, and genetic information scenarios, and do not consider multi-scenario situations. This is because sensitive personal information involves a wide range of scenarios, and it is impossible to simply collect all scenario data together for classification and grading. It is often necessary to select some scenario data for classification and grading based on the current reality. However, once the original scenario classification and grading is completed, if a new scenario appears, it is difficult to add the subsequent new scenario data to the original scenario classification and grading structure. Therefore, how to expand the new scenario data into the original scenario classification and grading structure through incremental methods is a challenge that needs to be solved first.

[0005] (2) In multi-scenario situations, the correlation between multiple scenarios needs to be considered. For example, if the same person appears in multiple scenarios, it is necessary to associate the attributes of the person appearing in different scenarios with the same person. Therefore, the correlation between multiple scenarios is also a challenge that needs to be addressed.

[0006] (3) With the rapid development of big data, artificial intelligence and the Internet of Things, how to effectively store massive heterogeneous data and ensure data retrieval efficiency is also a problem.

[0007] (4) Conventional classification and grading methods use relational databases to store the results (which is inevitable since they are not combined with knowledge graphs). Classification and grading are performed for a single scenario. If classification and grading have been completed in scenario A and then separately in scenario B, then there are two database tables to store the classification and grading results for the two scenarios. Now, it is necessary to combine the classification and grading results from scenarios A and B. Since scenarios A and B are different, the attributes of the entities considered are naturally different. Therefore, when storing the classification and grading results of the two scenarios at the same time, the database tables need to be redesigned and the data needs to be re-stored. Redesigning the database tables and re-integrating the data into a single table is time-consuming and labor-intensive. When there are many scenarios or a lot of personal data, the efficiency overhead will be even greater. However, the multi-knowledge graph fusion technology of knowledge graphs does not require changes to the storage structure and can be expanded. It is simple and efficient, so it has the scalability of classification and grading.

[0008] Regarding the first problem, existing methods for classifying and grading sensitive personal information often only consider implementation in a single scenario, neglecting the implementation of classification and grading data across multiple scenarios. Regarding the second problem, existing methods do not consider the correlation between multiple scenarios. Regarding the third and fourth problems, existing data is large in scale and growing rapidly, while many classification and grading schemes still use relational databases. Currently, relational databases struggle to handle the storage of massive amounts of unstructured data, and when the database is large, relational operations such as Cartesian products incur significant time overhead, drastically reducing data retrieval efficiency. In short, there is still no satisfactory solution to these four problems. Although some research has touched upon these issues to some extent, the considerations remain incomplete, making it difficult to provide a comprehensive solution.

[0009] Therefore, we propose a novel method for classifying and grading sensitive personal information using knowledge graphs. This method considers classification and grading across multiple real-world scenarios. When a new scenario emerges, it utilizes multi-knowledge graph fusion technology to achieve scenario fusion, thus realizing scalability across multiple scenarios. It leverages multi-knowledge graph fusion and knowledge reasoning techniques to address the correlation problem across multiple scenarios. Furthermore, this method is better suited for storing and retrieving classification and grading results for large-scale data, utilizing a knowledge graph database to store massive amounts of heterogeneous data and enabling efficient querying. Summary of the Invention

[0010] The purpose of this invention is to address the problem that existing methods for classifying and grading sensitive personal information often only consider the implementation of a single scenario and neglect the implementation of data classification and grading in multiple scenarios.

[0011] To solve the above-mentioned technical problems, the present invention adopts the following technical means:

[0012] A method for classifying and grading sensitive personal information using knowledge graphs includes the following steps:

[0013] Step 1: Extract entities and relationships from text data containing personal information to obtain a set of general entities and a set of relationships;

[0014] Step 2: First, construct a personal information knowledge graph based on the set of general entities and relationships;

[0015] Step 3: Extract features from each general entity in Step 1 to obtain a feature vector corresponding to each general entity;

[0016] Step 4: Cluster the feature vectors to obtain the clustering results for general entities;

[0017] Step 5: Classify each category of the clustering results in Step 4 into a security level classification to obtain sensitivity level information. Each level in the sensitivity level information corresponds to a level entity.

[0018] Step 6: Associate the sensitivity level information with the personal information knowledge graph to obtain a knowledge graph for classifying and grading personal sensitive information.

[0019] Step 7: When a new scenario arrives, incremental updates are performed using multi-knowledge graph fusion technology.

[0020] In the above technical solution, step 7 specifically includes the following steps:

[0021] Step 1: Using steps 1-6, construct a knowledge graph for classifying and grading personal sensitive information in the new scenario to obtain the knowledge graph for classifying and grading personal sensitive information in the new scenario.

[0022] Step 2: Perform entity alignment to find entities in the new scene that refer to the same entities as those in the original scene. These entities include general entities and hierarchical entities.

[0023] Step 3: Perform knowledge graph fusion.

[0024] If entity 'a' in the new scene already exists in the original scene, then find entity A directly associated with 'a' in the new scene. n Given an entity set A, select entities A that exist in the new scene but not in the original scene. ’ n Entity A ’ n Created within the existing scene and directly associated with entity a;

[0025] If entity b in the new scenario does not exist in the original scenario, then create a new entity b in the original scenario and find the entity B directly associated with b in the new scenario. n Given set B, select entities B that exist only in the new scene and not in the original scene. ’ n Entity B ’ n Created within the existing scenario and directly associated with entity b, it completed incremental updates, integrating information from the new scenario into the existing personal sensitive information classification and grading scenario.

[0026] The above technical solution also includes a knowledge reasoning step. After the incremental update is completed, knowledge reasoning technology is used to associate entities in the knowledge graph and determine whether the associated entities have attribute values. If there are no attribute values, the attribute values ​​of the associated entities are assigned to the entities without attribute values.

[0027] In the above technical solution, the knowledge reasoning steps specifically include the following steps:

[0028] A multi-classifier is trained using paths in the graph to infer the attribute values ​​of entities that lack attribute values. The specific steps include the following:

[0029] 1. Generate and select a set of paths for the relationships between a single entity and other entities, using depth-first search;

[0030] 2. Calculate the feature value of each path. This feature value is represented by a probability value, indicating the path from entity s. The probability of reaching entity node t;

[0031] 3. Based on the feature values ​​of the path, train a classifier. Once the classifier is trained, it can be used to infer the probability that two entities have a target relationship.

[0032] The classifier represents the starting node. and the end node The relationships that exist are:

[0033]

[0034] In the above formula, It is a weight of The path, It is a path The magnitude of the probability value, It is a connection and The set of all paths.

[0035] This algorithm is designed to determine whether a relationship exists between two entities. Using this algorithm, I can make inferences between Zhang San and existing professional entities, assuming that the classifier has already been trained.

[0036] Therefore, I can make a judgment:

[0037] A score is given to determine whether there is a connection between "Zhang San" and "doctor".

[0038] The question is scored based on whether there is a connection between "Zhang San" and "teacher".

[0039] Is there any connection between "Zhang San" and the "lawyer"...?

[0040] The final classifier infers Zhang San's most likely occupation as "xxx" based on the score.

[0041] Furthermore, when a small number of new entities are added to the knowledge graph, even if the classification of the entities is not yet known, this knowledge graph reasoning technology can be used to quickly reason about the relationship between entities and levels, and to quickly classify the entities.

[0042] In the above technical solution, when a number of new general entities less than the threshold are added to the knowledge graph, the knowledge graph reasoning technology is used to quickly infer the relationship between the newly added general entities and the existing level entities, and to quickly classify the entities, even when the classification of the general entities is not yet known.

[0043] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0044] (1) This is a multi-scenario problem involving sensitive personal information. Currently, most classification and grading schemes do not consider the scalability of combining multiple scenarios. When a new scenario arrives, how can it be incorporated into the existing classification and grading system? Our method uses knowledge graphs to classify and grade sensitive personal information, and retains the classification and grading results in the form of knowledge graphs. Knowledge graphs have high scalability. When sensitive personal information data from a new scenario arrives, there is no need to change the existing knowledge graph structure. By using the technology of combining multiple knowledge graphs, the fusion of new scenarios can be completed efficiently, making our scheme highly scalable.

[0045] (2) When personal sensitive information involves a large number of scenarios, there is a problem of correlation in multiple scenarios. For example, the same entity appears in multiple scenarios, and it is necessary to associate the relationship of the entity in each scenario with the same entity. Our classification and grading scheme can solve the problem of correlation in multiple scenarios. First, we use knowledge graph fusion technology to complete the fusion of knowledge graphs and associate entities in multiple scenarios. Then, we use knowledge reasoning technology to improve the correlation between entities.

[0046] (3) is the issue of data storage method and retrieval efficiency when storing large amounts of data. When using a relational database, the retrieval of the data may lead to huge relational operations. We store the results of classifying and grading personal sensitive information in the form of a knowledge graph. The result format of the knowledge graph is very suitable for using a graph database to store large-scale heterogeneous data. The efficient relational query performance of the graph database will greatly improve the retrieval efficiency.

[0047] (4) The knowledge reasoning part uses a graph-based reasoning algorithm. Reasoning can be used to complete attributes and quickly classify them, that is, to reason about potential attributes and quickly classify them. For the multi-scene correlation problem, a total personal sensitive information knowledge graph is first constructed, and incremental fusion is used to quickly associate existing attributes. Then, potential attributes are mined in the global personal sensitive information classification and classification knowledge graph. Completing the potential attributes can solve the multi-scene correlation problem. Attached Figure Description

[0048] Figure 1 This is a process for a method of classifying and grading sensitive personal information using knowledge graphs;

[0049] Figure 2 Example of a personal information knowledge graph;

[0050] Figure 3 Here is an example of entity clustering results;

[0051] Figure 4 Example of a knowledge graph for classifying and grading sensitive personal information. Detailed Implementation

[0052] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0053] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0054] We propose a method for classifying and grading sensitive personal information using knowledge graphs. The overall process of this method is as follows: Figure 1 As shown.

[0055] Suppose we have text data 1, 2, ..., m. Taking citizens' personal information as an example, the first step is to extract entities and relations from the text data, obtaining sets of entities and sets of relations. The second step is to construct a personal information knowledge graph based on the sets of entities and relations. Suppose we can obtain the entity "Zhang San" from the text, whose mobile phone number is "12345678910", whose ID number is "510000000000000107", whose address is "No. 1, Street C, District B, City A", whose bank card number is "1234-5678-1011-1213", and whose age is "30 years old". Then, when Zhang San is the center, the constructed personal information knowledge graph is as follows: Figure 2 As shown.

[0056] The third step is to extract features from the entities, obtaining a feature vector for each entity.

[0057] The fourth step is to cluster the feature vectors to obtain the clustering results of the entities (we use clustering as an example here, but supervised or semi-supervised learning techniques can also be used for category division). The results are as follows. Figure 3 As shown, we group entities with similar meanings together.

[0058] The fifth step is to determine the number of security levels, the basis for classifying them, and to conduct specific grading for each category. For the number of security levels and the basis for classifying them, relevant laws and regulations can be consulted. For example, the "Network Data Classification and Grading Guidelines" suggests the following grading rules:

[0059]

[0060] Table 1 Examples of Hierarchical Rules

[0061] When there are few categories, a purely manual approach can be used for classification. Classification can be tailored to the specific application scenario. For example, a reasonable classification case in this scenario would be as follows:

[0062]

[0063] Table 2. Case Studies of Hierarchy

[0064] When there are many categories, a combination of human and artificial intelligence can be used for classification. Specifically, human experience is used to classify some categories into levels, and then deep learning technology is used to predict the levels of unclassified categories by learning from existing level classifications.

[0065] The sixth step involves associating the sensitivity level information with the personal information knowledge graph to obtain a classification and grading knowledge graph for sensitive personal information. Specifically, this involves adding level entities to the personal information knowledge graph and, based on the grading results, directly associating each level entity with all entities belonging to that level. Figure 4 It is a reasonable way to display a knowledge graph centered on a person's name when the knowledge graph for classifying and grading sensitive personal information is completed.

[0066] Incremental update method when a new scenario arrives:

[0067] Once the knowledge graph for classifying and grading personal sensitive information in a single scenario is completed, we can incrementally update it using multi-knowledge graph fusion technology when new scenarios arise. One possible approach is as follows:

[0068] Step 1: We use our method to construct a knowledge graph for classifying and grading personal sensitive information in the new scenario, thus obtaining the knowledge graph for classifying and grading personal sensitive information in the new scenario.

[0069] Step 2: We perform entity alignment to find entities in the new scene that refer to the same entities as those in the original scene.

[0070] Step 3: We perform knowledge graph fusion. If entity 'a' in the new scenario already exists in the original scenario, we find entities directly associated with 'a' in the new scenario, select those entities that exist in the new scenario but not in the original scenario, create these entities in the original scenario, and directly associate them with 'a'. If entity 'b' in the new scenario does not exist in the original scenario, we create a new entity 'b' in the original scenario, find entities directly associated with 'b' in the new scenario, select those entities that exist in the new scenario but not in the original scenario, create these entities in the original scenario, and directly associate them with 'b'. Through these steps, we complete the incremental update, integrating information from the new scenario into the original personal sensitive information classification and grading scenario.

[0071] Solving correlation problems in multiple scenarios:

[0072] When sensitive personal information involves numerous scenarios, the issue of correlation across multiple scenarios arises. This means the same person appears in multiple scenarios, and it's necessary to associate their attributes from each scenario with this same person. This problem can be solved using multi-knowledge graph fusion technology and knowledge reasoning technology. For example, taking Zhang San as an example, a specific implementation method is as follows:

[0073] Step 1: We create a knowledge graph of sensitive personal information for each scenario, resulting in Knowledge Graph 1 of Zhang San's Sensitive Information for Scenario 1, Knowledge Graph 2 of Zhang San's Sensitive Information for Scenario 2, and so on.

[0074] Step 2: We integrate the knowledge graphs incrementally to obtain the total personal sensitive information knowledge graph.

[0075] Step 3: We use knowledge reasoning technology on the overall personal sensitive information knowledge graph to improve the relationships between entities.

[0076] Solving correlation problems in multiple scenarios:

[0077] When sensitive personal information involves numerous scenarios, the issue of correlation across multiple scenarios arises. This means the same person appears in multiple scenarios, and it's necessary to associate their attributes from each scenario with this same person. This problem can be solved using multi-knowledge graph fusion technology and knowledge reasoning technology. For example, taking Zhang San as an example, a specific implementation method is as follows:

[0078] Step 1: We create a knowledge graph of sensitive personal information for each scenario, resulting in Knowledge Graph 1 of Zhang San's Sensitive Information for Scenario 1, Knowledge Graph 2 of Zhang San's Sensitive Information for Scenario 2, and so on.

[0079] Step 2: We perform incremental fusion of the knowledge graph to obtain the total personal sensitive information knowledge graph, thus obtaining the known attributes of Zhang San when he appears in different scenarios.

[0080] Step 3: We infer Zhang San's potential attributes when he appears in different scenarios. We use knowledge reasoning techniques on a global knowledge graph for classifying and grading sensitive personal information to achieve this. For a specific example, suppose Zhang San's occupational attribute value is empty, while others have occupational attribute values. We can use paths to train a multi-classifier in the graph to infer Zhang San's occupation.

[0081] Specifically, we will train a score to represent the starting node. and the end node Is there some kind of relationship between them?

[0082]

[0083] In the above formula, It is a weight of The path, It is a path The magnitude of the probability value, It is a connection and The set of all paths;

[0084] 1. Generate and select a set of paths for the relationships between a single entity and other entities, using depth-first search.

[0085] 2. Calculate the feature value of each path. This feature value is represented by a probability value, indicating the path from entity s. The probability of reaching entity node t;

[0086] 3. Train a classifier for the target relationship based on the feature values ​​of the training examples. Once the classifier is trained, it can be used to infer whether a target relationship exists between two entities.

[0087] Therefore, we can infer whether there is a relationship between Zhang San and a certain professional entity, and thus infer Zhang San's profession.

[0088] This algorithm is designed to determine whether a relationship exists between two entities. Using this algorithm, I can make inferences between Zhang San and existing professional entities, assuming that the classifier has already been trained.

[0089] Therefore, I can make a judgment:

[0090] A score is given to determine whether there is a connection between "Zhang San" and "doctor".

[0091] The question is scored based on whether there is a connection between "Zhang San" and "teacher".

[0092] Is there any connection between "Zhang San" and the "lawyer"...?

[0093] The final classifier infers Zhang San's most likely occupation as "xxx" based on the score.

[0094] Furthermore, when a small number of new entities are added to the knowledge graph, even if the classification of the entities is not yet known, this knowledge graph reasoning technology can be used to quickly reason about the relationship between entities and levels, and to quickly classify the entities.

[0095] In the above technical solution, when a number of new general entities less than the threshold are added to the knowledge graph, the knowledge graph reasoning technology is used to quickly infer the relationship between the newly added general entities and the existing level entities, and to quickly classify the entities, even when the classification of the general entities is not yet known.

[0096] In simple terms, the process begins by training a classifier to determine the relationship between entities and levels. Then, a new entity is input into the classifier to obtain the level most closely associated with it. For example, determining whether a "fingerprint" is associated with "Level 1" yields a score; determining whether it's associated with "Level 2" yields a score; determining whether it's associated with "Level 3" yields a score; determining whether it's associated with "Level 4" yields a score; and finally, the "fingerprint" is associated with the level that has the highest score.

[0097] The present invention includes the following references:

[0098] [1] National People's Congress of China. Personal Information Protection Law of the People's Republic of China [EB / OL]. http: / / www.npc.gov.cn / npc / c30834 / 202108 / a8c4e3672c74491a80b53a172bb753fe.shtml, 2021-08-20

[0099] [2] National People's Congress of China. Data Security Law of the People's Republic of China [EB / OL]. http: / / www.npc.gov.cn / npc / c30834 / 202106 / 7c9af12f51334a73b56d7938f99a788a.shtml, 2021-06-10

[0100] [3] China.com.cn. Four departments jointly issued a document clarifying the scope of necessary personal information for 39 common types of apps [EB / OL]. https: / / baijiahao.baidu.com / s?id=1694910769977570489&wfr=spider&for=pc, 2021-03-22

[0101] [4] National Information Security Standardization Technical Committee. 14 National Cybersecurity Standards Approved and Released [EB / OL]. 2022-10-19

[0102] [5] He Weiqun. Personal information protection based on classification and grading [J]. Information Security and Communication Confidentiality, 2021(10):107-114.

[0103] [6] Li Jing, Qi Guoqiang, Yu Gang. Research and design of intelligent classification and grading system for children's medical data [J]. Electronic Technology and Software Engineering, 2022(15):228-231.

[0104] [7] Lu Hong. A method for classifying and grading urban data based on deep learning clustering algorithm [J]. Industrial Technology Innovation, 2021, 08(04):73-78. DOI:10.14103 / j.issn.2095-8412.2021.08.012.

[0105] The above-mentioned literature [5] is based on classification and grading personal information protection: the scheme constructs a classification and grading related personal information protection technical architecture, in which the classification and grading part is based on the micro data of personal information, specifically including each (row) record of micro data, each field (column) in the micro data record, and the data table composed of micro data as the classification and grading object of personal information. Compared with [5], our method classifies and grades the objects of classification and grading using entity extraction technology. We do not need to build a structured information database in advance. Obviously, extracting a structured information database from the data source will be more complicated and costly. Our method does not take the structured information data source as a premise, making our method more flexible and more versatile. Although the method in [5] takes into account the situation of multiple data scenarios, it does not consider how to achieve the expansion of new classification and grading scenarios through incremental methods. When a new scenario comes, [5] needs to redesign the data table storage method and then add new data to all the previous data to reclassify and grade the data. The overhead of expanding new scenarios is huge, while our method can efficiently complete the fusion of new scenarios by using the technology of combining multiple knowledge graphs; [5] does not consider the correlation between multiple scenarios, while our solution can combine the technology of multi-knowledge graph fusion and knowledge reasoning to discover the correlation between multiple scenario data; when the data scale is huge, the solution in [5] will be difficult to implement by relying solely on manual labeling, while our method also provides a manual + intelligent classification and grading approach; the solution in [5] does not consider the storage and retrieval efficiency of the results after classification and grading, while the knowledge graph proposed in our solution provides a good solution for the storage and query of these classification and grading results.

[0106] Reference [6] Research and Design of Intelligent Classification and Grading System for Pediatric Medical Data: This scheme designs an intelligent classification and grading method. Through the process of rule definition and data resource catalog construction, an intelligent classification and grading system for pediatric medical data is developed. Compared with [6], [6] requires formatting of pediatric data before classification and grading, while our scheme does not require formatting preprocessing, making it more flexible and efficient. [6] still does not consider the incremental situation of multiple data scenarios. For example, when [6] collects pediatric hospital information for three scenarios and completes classification and grading, if it is required to add pediatric hospital information for a fourth scenario, it will cause the scheme in [6] to reclassify and grade the entire process. Our scheme is more scalable for multiple scenarios and can efficiently complete the fusion of new scenarios by using the technology of combining multiple knowledge graphs. The scheme in [6] also does not consider the storage and retrieval efficiency of the results after classification and grading. The knowledge graph proposed in our scheme provides a good solution for the storage and query of these classification and grading results.

[0107] Reference [7] proposes a method for classifying and grading urban data based on a deep learning clustering algorithm. Compared with [7], the method in [7] also has the problem of difficulty in incremental scene addition after the model is launched. Adding data from new scenes means retraining the model, which has a high cost. The method in [7] also does not consider the efficiency of storing and retrieving the results after classification and grading.

Claims

1. A method for classifying and grading sensitive personal information using knowledge graphs, characterized in that... This includes the following steps: Step 1: Extract entities and relationships from text data containing personal information to obtain a set of general entities and a set of relationships; Step 2: First, construct a personal information knowledge graph based on the set of general entities and relationships; Step 3: Extract features from each general entity in Step 1 to obtain a feature vector corresponding to each general entity; Step 4: Cluster the feature vectors to obtain the clustering results for general entities; Step 5: Classify each category of the clustering results in Step 4 into a security level classification to obtain sensitivity level information. Each level in the sensitivity level information corresponds to a level entity. Step 6: Associate the sensitivity level information with the personal information knowledge graph to obtain a knowledge graph for classifying and grading personal sensitive information; Step 7: When a new scenario arrives, incremental updates are performed using multi-knowledge graph fusion technology; Step 7 specifically includes the following steps: Step 7.1: Using steps 1-6, construct a knowledge graph for classifying and grading personal sensitive information in the new scenario to obtain the knowledge graph for classifying and grading personal sensitive information in the new scenario. Step 7.2: Perform entity alignment to find entities in the new scene that refer to the same entities as those in the original scene. These entities include general entities and hierarchical entities. Step 7.3: Perform knowledge graph fusion. If entity a in the new scene already exists in the original scene, find the entity An that is directly associated with a in the new scene to obtain entity set A. Select entity A'n that exists in the new scene but not in the original scene from entity set A, create entity A'n in the original scene and directly associate it with entity a. If entity b in the new scenario does not exist in the original scenario, a new entity b is created in the original scenario, and an entity Bn directly associated with b is found in the new scenario to obtain a set B. Entities B'n that exist only in the new scenario and not in the original scenario are selected from the entity set B, and entity B'n is created in the original scenario and directly associated with entity b, thus completing the incremental update and integrating the information of the new scenario into the original scenario of personal sensitive information classification and grading.

2. The method for classifying and grading personal sensitive information using knowledge graphs according to claim 1, characterized in that... It also includes a knowledge reasoning step. After the incremental update is completed, knowledge reasoning technology is used to associate entities in the knowledge graph and determine whether the associated entities have attribute values. If they do not have attribute values, the attribute values ​​of the associated entities are assigned to the entities without attribute values.

3. The method for classifying and grading personal sensitive information using knowledge graphs according to claim 2, characterized in that... The knowledge reasoning process specifically includes the following steps: A multi-classifier is trained using paths in the graph to infer the attribute values ​​of entities lacking attribute values. The specific steps include: Step 4.

1. Generate and select a set of paths for the relationships between a single entity and other entities, using depth-first search; Step 4.

2. Calculate the feature value of each path. This feature value is represented by a probability value, indicating the path from entity s. The probability of reaching entity node t; Step 4.

3. Train a classifier based on the feature values ​​of the path. Once the classifier is trained, it can be used to infer the probability that two entities have a target relationship. The classifier represents the starting node. and the end node The relationships that exist are: In the above formula, It is a weight of The path, It is a path The magnitude of the probability value, It is a connection and The set of all paths.

4. The method for classifying and grading personal sensitive information using knowledge graphs according to claim 1, characterized in that, When a new general entity is added to the knowledge graph in a number less than the threshold, the knowledge graph reasoning technology is used to quickly infer the relationship between the newly added general entity and the existing level entities, and to quickly classify the entities, even if the classification of the general entity is not yet known.

Images