Classification and grading identification method and system for threat data

By building refined network security data dictionary and Bert models and other technologies, the complexity and privacy and security issues of data classification and grading in the big data era are solved, and the precise classification and grading of network traffic threat data is realized, which improves processing efficiency and accuracy , ensuring the privacy, security and compliance of data.

CN118940145AInactive Publication Date: 2024-11-12INST OF SOFTWARE - CHINESE ACAD OF SCI
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410943933.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the era of big data, automated data classification and grading face difficulties in integrating multimodal information, information redundancy and noise interference. Concept drift and complex contexts make it difficult for the system to follow up on real-time changes. The high labeling costs and privacy security issues are highlighted. Advanced automated classification and grading technology to quickly and accurately extract valuable intelligence from massive data while ensuring privacy, security and compliance of information.

Method used

By building a refined network security data dictionary, combining feature extraction, semantic matching, similarity calculation and sensitivity scoring, using Bert model and full link layer technologies, we can achieve accurate classification and grading of network traffic threat data, including building multiple categories Data dictionary, word participle, word matching, feature vector splicing, information fusion and sensitivity score calculation determine the importance and level of data.

Benefits of technology

It realizes accurate classification and grading of network traffic threat data, improves the efficiency and accuracy of data processing, reduces the risks of manual intervention and errors, ensures the consistency and reliability of data processing results, and enhances the accuracy and privacy of grading. Safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940145A_ABST
    Figure CN118940145A_ABST
Patent Text Reader

Abstract

The invention discloses a classification and grading identification method and system for threat data, and belongs to the field of natural language processing. According to the method, a refined network security data dictionary is constructed, dictionary word matching is carried out on obtained network traffic data to determine hit texts, then the hit text with the highest similarity is obtained through a Bert model and similarity calculation, and then the splicing features are subjected to information fusion and feature value index operation to determine categories. Then, feature extraction and observation attribute matching are carried out to determine the importance level, and finally, the specific level is determined through sensitivity score calculation. According to the method, accurate characterization and accurate classification and grading of the threat data of the network traffic are realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to a method and system for classifying and grading threat data. Background Art

[0002] In the early days, data classification and grading technology mainly relied on manual and rule-based methods. Data managers or experts manually classified and graded data according to the nature, source and purpose of the data. Although this method is intuitive, it is inefficient and is unable to handle large-scale data sets. With the rapid development and widespread application of information technology, a large amount of sensitive data is generated and stored, such as personal privacy data and commercial confidential data. Data classification and grading, as an important data management technology, aims to classify and label data according to its sensitivity in order to achieve data protection and access control. With the gradual development of automation and machine learning technology, data classification and grading have begun to be automated. By training models and algorithms, computers can intelligently classify and grade data. Data classification is the basis for data classification and labeling. This automated method greatly improves the efficiency and accuracy of classification and grading, and reduces the need for manual intervention.

[0003] Data classification technology based on big data aims to realize automatic classification of data, including rule-based data classification methods and statistics-based data classification methods.

[0004] Rule-based data classification methods mainly include FOIL algorithm, decision tree algorithm, association classification method, etc. FOIL algorithm effectively reduces redundant rules, but each training set is only covered by one rule, so the accuracy of classification is not high, especially when the training set is small. The decision tree algorithm uses a tree structure to build a classification model, providing a method to make corresponding decisions according to different situations. The disadvantage of the decision tree is that for some data sets, when there are many instances in the data set, the resulting decision tree is very large and the classification model is more complex, so the decision tree needs to be simplified. The association classification method has a higher accuracy than the decision tree algorithm classification. Its focus is to identify the connection and interaction between various variables in the data set. Its goal is to find the relationship and pattern between attributes in order to predict future events or classify new events. Association classification can be used to discover useful patterns to help enterprises and organizations better understand their data, make data-driven decisions, and improve operations.

[0005] The statistical data classification methods mainly include the following algorithms: k-nearest neighbor method (KNN), naive Bayes method, Bayesian network, support vector machine (Support Vector Machine), SOM algorithm, among which SVM classifier and KNN classifier are currently the two classifiers with the best classification performance.

[0006] Data classification technology based on big data aims to realize automatic classification of data, which mainly includes content-based classification method, context-based classification method and rule-based classification method.

[0007] The content-based classification method is to classify and mark the data according to its actual content. For example, for text data, technologies such as keyword matching, text analysis, and natural language processing can be used to identify and mark sensitive words or sensitive information. For image data, technologies such as image recognition and feature extraction can be used to identify and classify images containing sensitive content.

[0008] The context-based classification method considers the contextual information of data usage. This contextual information can include factors such as the source of the data, the identity of the visitor, and the usage environment. For example, within an enterprise, data can be classified and labeled according to the position and authority of employees to achieve different levels of access control.

[0009] The rule-based classification method uses predefined rules and policies to classify and label data. These rules can be internal policy and regulatory requirements of the organization or industry standards. For example, personal data is classified and labeled according to the privacy regulations of a certain region to ensure data compliance and privacy protection.

[0010] In the era of big data, faced with the influx of multi-source and multi-type data from all parties, automated classification and grading faces huge challenges. Data heterogeneity makes it difficult to integrate multimodal information, information redundancy and noise interfere with accuracy, and concept drift and complex context make it difficult for the system to follow up on real-time changes. The need to process large-scale data makes the efficiency of the algorithm a bottleneck, while the high annotation cost and privacy and security issues become more prominent. For complex and sensitive situations, advanced automated classification and grading technology is needed to quickly and accurately extract valuable intelligence from massive data while ensuring the privacy, security and compliance of information. Summary of the invention

[0011] The purpose of the present invention is to provide an efficient and accurate threat data classification and grading identification method, which can achieve accurate characterization and classification and grading of threat data of network traffic by constructing a refined data level of network security data dictionary, through feature extraction, semantic matching, similarity calculation and sensitivity scoring.

[0012] To achieve the above object, the present invention adopts the following technical solutions:

[0013] A method for classifying and grading threat data includes the following steps:

[0014] Construct a cybersecurity data dictionary containing multiple categories, determine the observation attributes of each category, and classify the importance level of the data and the level within each level;

[0015] Obtain network traffic data and perform word segmentation to obtain input text;

[0016] According to the matching degree between the words in the input text and the words in the network security data dictionary, the matched words in the dictionary are found and concatenated to form multiple hit texts;

[0017] Use the Bert model to convert the input text and hit text into feature vectors, find the hit text with the highest similarity to the input text, and then concatenate the feature vectors of the input text and the hit text with the highest similarity;

[0018] The concatenated vector is input into the full-link layer for information fusion, and then the vector feature value is exponentially calculated through the Softmax layer, and then normalized to determine the category to which the input text belongs;

[0019] Extract features from the input text and determine the importance level of the input text based on the inclusion relationship or similarity between the extracted features and the observed attributes of the category to which they belong;

[0020] The sensitivity score of the feature is calculated based on the relationship between the extracted feature and the data in the network security data dictionary, the total score of the input text is determined by weighted average, and the level of the input text in the level is determined based on the total score.

[0021] Furthermore, the cybersecurity data dictionary contains three categories: source and type, organization and agency name, and key vocabulary in key areas.

[0022] Furthermore, based on the degree of matching between the words in the input text and the words in the network security data dictionary, the matched words in the dictionary are found, which means that if the words in the input text are exactly the same as the words in the network security data dictionary or the cosine similarity between the two is equal to or exceeds a similarity threshold, then the words in the network security data dictionary are considered to be matched words.

[0023] Furthermore, the concatenated vector is input into the full-link layer for information fusion, which means that the concatenated vector is input into the full-link layer for linear combination and transformation, and the features in the concatenated vector are weighted summed to achieve information fusion of each feature.

[0024] Furthermore, the importance levels of data include core data, important data and general data, and each level includes multiple grades.

[0025] Furthermore, the step of determining the importance level of the input text based on the inclusion relationship or similarity between the extracted features and the observed attributes of the category to which they belong includes: judging whether the extracted features contain any observed attributes of the category to which they belong, and if so, determining that the input text is core data; if not, calculating the cosine similarity between the extracted features and all observed attributes of the category to which they belong, and if the maximum cosine similarity is equal to or exceeds a similarity threshold, determining it to be important data, otherwise determining it to be general data.

[0026] Furthermore, the step of calculating the sensitivity score of the feature based on the relationship between the extracted feature and the data in the network security data dictionary includes: if the extracted feature is included in the network security data dictionary, the sensitivity score is 1; otherwise, the Pearson correlation coefficient between the extracted feature and each data in the network security data dictionary is calculated, and the maximum value of the Pearson correlation coefficient is taken as the sensitivity score.

[0027] Furthermore, the step of determining the level in the level of the input text according to the total score includes: determining the level division rules, each level division is based on the range defined by the size threshold as the standard, and determining the level of the input text according to the specific range where the total score is located.

[0028] Furthermore, a level division rule is as follows: level one, 90<total score ≤ 100; level two, 80<total score ≤ 90; level three, 70<total score ≤ 80; level four, 60<total score ≤ 70; level five, 50<total score ≤ 60; level six, 40<total score ≤ 50; level seven, 30<total score ≤ 40; level eight, 20<total score ≤ 30; level nine, 10<total score ≤ 20; level ten, 0<total score ≤ 10.

[0029] A classification and grading identification system for threat data, comprising:

[0030] The data classification and recognition module is used to find the matched hit words in the dictionary according to the matching degree between the words in the input text and the words in the network security data dictionary, and splice them into multiple hit texts; use the Bert model to convert the input text and the hit text into feature vectors, find the hit text with the highest similarity to the input text, and then splice the input text with the hit text with the highest similarity into feature vectors; input the spliced ​​vector into the full link layer for information fusion, and then perform exponential operation on the vector feature value through the Softmax layer, and then determine the category to which the input text belongs after normalization;

[0031] The data classification and recognition module is used to extract features from the input text and determine the importance level of the input text based on the inclusion relationship or similarity between the extracted features and the observed attributes of the category to which they belong; calculate the sensitivity score of the feature based on the relationship between the extracted features and the data in the network security data dictionary, determine the total score of the input text by weighted average, and determine the level in the level of the input text based on the total score.

[0032] The present invention has the advantages that:

[0033] 1. The present invention provides a basis for accurate identification of threat data by constructing a sophisticated network security data dictionary, determining the observed attributes of categories, and dividing the importance levels and grades.

[0034] 2. When classifying data, the present invention realizes a complete processing flow from raw network traffic data to structured data, and then to final classification and grading through rule matching, Bert model feature extraction, cosine similarity calculation, etc. The Bert model is used to deeply understand the text semantics, and the classification ability of the model is further enhanced by introducing the fully connected layer and the Softmax layer, making the classification results more accurate, reducing manual intervention and error risks, improving the efficiency and accuracy of data processing, and ensuring the consistency and reliability of data processing results.

[0035] 2. When classifying data, the present invention determines the importance level by the matching relationship or similarity with the observed attributes, and can accurately realize grade identification through sensitivity scoring. The Pearson correlation coefficient algorithm is used to fully consider the correlation between data, thereby improving the accuracy of classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The figure is an overall flow chart of the method for classifying and grading threat data according to an embodiment of the present invention.

[0037] Figure 2 A schematic diagram showing how to determine the attributes of an observation based on its category.

[0038] Figure 3 The present invention is a flowchart of data classification and identification according to an embodiment of the present invention.

[0039] Figure 4 Example graph of obtaining hit text based on logical hit rules and similarity matching.

[0040] Figure 5 This is an example diagram of data classification and recognition based on the Bert model.

[0041] Figure 6 The flowchart of data classification identification according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below.

[0043] The embodiment of the present invention specifically discloses a classification and classification identification method for threat data, which is described below in terms of three parts: dictionary construction, data classification identification, and data classification identification. The overall process is as follows: Figure 1 shown.

[0044] 1. Dictionary construction

[0045] (1) Construct a cybersecurity data dictionary, which includes the following three categories:

[0046] a) Sources and types, including signal-based cybersecurity data, personnel-based cybersecurity data, open source cybersecurity data, image-based cybersecurity data, measurement and signature-based cybersecurity data, financial cybersecurity data, technical cybersecurity data, and threat-based cybersecurity data.

[0047] b) Organizational structure and institution name, including technical security team, information security management department, security operation center (SOC), cybersecurity institutions and institutions, and other cybersecurity-related departments.

[0048] c) Key words in key areas. Key areas include economy, security and other areas. Each area contains some key words, such as organization name, legal representative name, product name, etc.

[0049] (2) Determine the observation attributes of each category based on the characteristics of the categories in the dictionary, such as Figure 2 shown.

[0050] (3) Constructing the level and grade of data in the dictionary. The data classification needs to be based on a certain dimension, otherwise the levels between different dimensions are not comparable. Therefore, the present invention combines the two dimensions of data category and data importance to achieve data level determination.

[0051] a) Determine the importance of the data. According to the affected objects and the degree of impact, the importance of the data is divided into three levels: core data, important data and general data. A well-known classification rule is shown in Table 1 below.

[0052] Table 1 Data classification rules

[0053]

[0054] b) For the above three levels, each level can be divided into multiple grades. The following Table 2 is an example of division into 10 grades.

[0055] Table 2 Data level division rules

[0056] Core Data Level 1, Level 2, Level 3...Level 10 Important data Level 1, Level 2, Level 3...Level 10 General Data Level 1, Level 2, Level 3...Level 10

[0057] Through the above operations, a “category-observation attribute-level” system was constructed.

[0058] 2. Data classification and identification, the process is as follows Figure 3 shown.

[0059] (1) Obtaining hit text based on logical hit rules and similarity matching based on the network security data dictionary:

[0060] a) Based on the intercepted network traffic data, the word segmentation technology is used to divide the continuous text into meaningful word sets as the subsequent input text;

[0061] b) The word set is matched with the words in the network security data dictionary. When a word in the input word set is exactly the same as a word in the dictionary, it is considered that an exact match has occurred; when the input word is not exactly the same as a word in the dictionary, the cosine similarity algorithm is used to evaluate the similarity between them. When the cosine similarity between the two words exceeds a similarity threshold (the similarity threshold is set to 0.8 in this embodiment), they are considered to be similar; the words in the dictionary that are exactly matched or similar are regarded as hit words.

[0062] c) Randomly combine the hit words and concatenate them into text content to obtain multiple hit texts.

[0063] Figure 4 An example of obtaining hit text based on logical hit rules and similarity matching is shown.

[0064] (2) Classification and recognition based on the Bert model:

[0065] a) Use the Bert model to convert the input text and the hit text into feature vector representations respectively, and then use the cosine similarity algorithm to calculate the similarity between the two, and mine the hit text with the highest similarity to the input text. Finally, compare the feature vector Vec of the input text with the feature vector Vec′ of the hit text with the highest similarity. max Perform vector concatenation as input for subsequent algorithms.

[0066] b) Add a fully connected layer and a Softmax layer based on the Bert model. Input the concatenated vector into the fully connected layer for linear combination and transformation. By weighted summing the features in the concatenated vector, the information of different features is fused and a new feature representation is generated. Then the generated new feature representation is input into the Softmax layer, and each element (i.e., eigenvalue) in the representation feature is exponentially operated. All exponential operation results are normalized to obtain a probability distribution, and the classification result of the input text is determined based on the probability distribution.

[0067] Figure 5 An example of data classification and recognition based on the Bert model is shown.

[0068] 3. Data classification and identification, the process is as follows Figure 6 shown.

[0069] 1) Perform feature extraction on the input text to extract feature vector information such as names, organizations, structures, etc.

[0070] 2) Determine the importance level of the input text based on the extracted features, and the determination method is as follows: when the extracted features contain any observed attribute of the category to which they belong, the input text is determined as core data; if not, the cosine similarity algorithm is used to calculate the similarity value between the extracted features and all observed attributes of the category. When the maximum similarity value is equal to or greater than a similarity threshold (the similarity threshold is set to 0.95 in this embodiment), the input text is determined as important data, otherwise the input text is determined as general data.

[0071] 3) Determine the specific level of the input text based on the extracted features.

[0072] a) Calculate the sensitivity score of the extracted feature information such as names, organizations or structures: When the extracted feature is included in the network security data dictionary, the sensitivity score is 1; otherwise, calculate the Pearson correlation coefficient between the extracted feature and the data in the network security data dictionary, and the sensitivity score takes the maximum value of the Pearson correlation coefficient. Finally, calculate the sensitivity scores of all features, perform weighted average, and obtain the total score of the input text (0≤total score≤100).

[0073] b) Determine which level the text belongs to based on the total score of the input text and the score relationship of different levels. The division rules of different levels are to determine the score range based on the size threshold. In this embodiment, the division rules are as follows: the total score range corresponding to the first level is 90 < total score ≤ 100; the total score range corresponding to the second level is 80 < total score ≤ 90; the total score range corresponding to the third level is 70 < total score ≤ 80; the total score range corresponding to the fourth level is 60 < total score ≤ 70; the total score range corresponding to the fifth level is 50 < total score ≤ 60; the total score range corresponding to the sixth level is 40 < total score ≤ 50; the total score range corresponding to the seventh level is 30 < total score ≤ 40; the total score range corresponding to the eighth level is 20 < total score ≤ 30; the total score range corresponding to the ninth level is 10 < total score ≤ 20; the total score range corresponding to the tenth level is 0 < total score ≤ 10.

[0074] Through the processing of the above steps, the threat data in the acquired network traffic can be classified and graded.

[0075] The above examples are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A classification and grading method for identifying threat data, characterized in that: The following steps are involved: Construct a cybersecurity data dictionary containing multiple categories, determine the observation attributes of each category, and classify the importance level of the data and the level within each level; Obtain network traffic data and perform word segmentation to obtain input text; According to the matching degree between the words in the input text and the words in the network security data dictionary, the matched words in the dictionary are found and concatenated to form multiple hit texts; Use the Bert model to convert the input text and hit text into feature vectors, find the hit text with the highest similarity to the input text, and then concatenate the feature vectors of the input text and the hit text with the highest similarity; The concatenated vector is input into the full-link layer for information fusion, and then the vector feature value is exponentially calculated through the Softmax layer, and then normalized to determine the category to which the input text belongs; Extract features from the input text and determine the importance level of the input text based on the inclusion relationship or similarity between the extracted features and the observed attributes of the category to which they belong; The sensitivity score of the feature is calculated based on the relationship between the extracted feature and the data in the network security data dictionary, the total score of the input text is determined by weighted average, and the level of the input text in the level is determined based on the total score.

2. The method according to claim 1, characterized in that The Cybersecurity Data Dictionary contains three categories: source and type, organization and agency name, and key vocabulary in key areas.

3. The method according to claim 1, characterized in that According to the matching degree between the words in the input text and the words in the network security data dictionary, the matched words in the dictionary are found, which means that if the words in the input text are exactly the same as the words in the network security data dictionary or the cosine similarity between the two is equal to or exceeds a similarity threshold, the words in the network security data dictionary are considered to be matched words.

4. The method according to claim 1, characterized in that Inputting the concatenated vector into the full-link layer for information fusion means inputting the concatenated vector into the full-link layer for linear combination and transformation, performing weighted summation on the features in the concatenated vector, and realizing information fusion of each feature.

5. The method according to claim 1, characterized in that The importance levels of data include core data, important data and general data, and each level contains multiple grades.

6. The method according to claim 5, characterized in that The step of determining the importance level of the input text according to the inclusion relationship or similarity between the extracted features and the observed attributes of the category to which they belong includes: judging whether the extracted features contain any observed attributes of the category to which they belong, and if so, determining that the input text is core data; if not, calculating the cosine similarity between the extracted features and all observed attributes of the category to which they belong, and if the maximum cosine similarity is equal to or exceeds a similarity threshold, determining that it is important data, otherwise determining that it is general data.

7. The method according to claim 1, characterized in that The step of calculating the sensitivity score of the feature according to the relationship between the extracted feature and the data in the network security data dictionary includes: if the extracted feature is included in the network security data dictionary, the sensitivity score is 1; otherwise, the Pearson correlation coefficient between the extracted feature and each data in the network security data dictionary is calculated, and the maximum value of the Pearson correlation coefficient is taken as the sensitivity score.

8. The method according to claim 1, characterized in that The step of determining the level in the level of the input text according to the total score includes: determining the level division rules, each level division is based on the range defined by the size threshold as the standard, and determining the level of the input text according to the specific range where the total score is located.

9. The method according to claim 8, characterized in that The classification rules of one level are as follows: Level 1, 90 < total score ≤ 100; Level 2, 80 < total score ≤ 90; Level 3, 70<total score ≤ 80; Level 4, 60<total score ≤ 70; Level 5, 50<total score ≤ 60; Level 6, 40<total score ≤ 50; Level 7, 30<total score ≤ 40; Level 8, 20<total score ≤ 30; Level 9, 10<total score ≤ 20; Level X, 0<total score ≤ 10.

10. A classification and grading identification system for threat data, implementing the method according to any one of claims 1 to 9, characterized in that: include: The data classification and recognition module is used to find the matched hit words in the dictionary according to the matching degree between the words in the input text and the words in the network security data dictionary, and splice them into multiple hit texts; use the Bert model to convert the input text and the hit text into feature vectors, find the hit text with the highest similarity to the input text, and then splice the input text with the hit text with the highest similarity into feature vectors; input the spliced ​​vector into the full link layer for information fusion, and then perform exponential operation on the vector feature value through the Softmax layer, and then determine the category to which the input text belongs after normalization; The data classification and recognition module is used to extract features from the input text and determine the importance level of the input text based on the inclusion relationship or similarity between the extracted features and the observed attributes of the category to which they belong; calculate the sensitivity score of the feature based on the relationship between the extracted features and the data in the network security data dictionary, determine the total score of the input text by weighted average, and determine the level in the level of the input text based on the total score.

Citation Information

Patent Citations

  • Civil aviation microblog security public opinion sentiment analysis method based on deep learning

    CN108536801A

  • Threat knowledge extraction method and system combining rule matching and pre-training language model

    CN116707928A

  • Data sensitivity determination method and device, equipment and storage medium

    CN117009611A

  • Network security domain named entity identification method and system based on auxiliary vector

    CN117010382A