Project data security information grading processing method

Abstract

The application discloses a project data security information grading processing method, relates to the technical field of data security information grading, and extracts data security judgment rules by using a data compliance library to form a data security word library; an automatic grading judgment model is established by using natural language processing technology based on data security word library and semantic context analysis; the model is used for format and semantic recognition of database data, and based on the recognition result, the safety, compliance and sensitivity of data content are judged; the correlation degree of data entities and the security word library is calculated, the sensitivity degree is automatically recognized, and the data entities are divided into security levels based on the security features of the data entities. The application improves the intelligent level of data security grading, reduces the artificial workload, and improves the data processing efficiency; the natural language processing and deep learning technology are applied to improve the understanding and processing capacity for unstructured data.

Description

Technical Field

[0001] This invention relates to the field of data security information classification technology, and in particular to a method for classifying and processing project data security information. Background Technology

[0002] In the early days, data security classification relied mainly on manual methods, which was inefficient. In the 2000s, semi-automated technologies such as rule engines were introduced. In the 2010s, with the development of machine learning and NLP, technologies such as natural language processing and deep learning began to be applied to achieve intelligent security classification. Currently, semantic analysis technologies based on NLP and knowledge graphs are relatively mature and can effectively extract semantic features from text. Security association judgment models based on machine learning and deep learning are also becoming increasingly sophisticated.

[0003] Existing knowledge graph-based methods are mainly limited to single-hop reasoning for security associations, and cannot perform complex multi-hop reasoning. They also have a weak ability to detect implicit associations. Furthermore, because current methods rely on data-driven approaches, they lack a clear explanation of the security classification decision-making process, which to some extent reduces users' trust. Summary of the Invention

[0004] In view of the problems existing in the above-mentioned data security classification, this invention is proposed.

[0005] Therefore, the problem to be solved by this invention is how to provide a method to improve the intelligence level of data security classification, reduce manual workload, and improve data processing efficiency.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a method for classifying and processing project data security information, including extracting data security judgment rules from a data compliance library to form a data security lexicon; using natural language processing technology, based on the data security lexicon and semantic context analysis, establishing an automatic classification judgment model; performing format and semantic recognition on database data, and matching security roots based on the recognition results to determine the degree of security association; calculating the correlation between data entities and the security lexicon, automatically identifying the sensitivity, and classifying data entities into security levels based on their security characteristics.

[0008] As a preferred embodiment of the data security information hierarchical processing method described in this invention, the process of constructing the data security thesaurus includes the following steps: collecting relevant industry data compliance databases, using text analysis of the compliance database, and extracting data security and risk control keywords from compliance regulations; processing the keywords, extracting word roots, merging synonyms, and constructing a data security thesaurus; summarizing the keywords in the thesaurus for data security judgment; constructing a security word vector space, calculating the security relevance of words, and filtering low-relevance words.

[0009] As a preferred embodiment of the hierarchical processing method for project data security information described in this invention, the following steps are included: constructing a security word vector space: using Word2Vec technology to train a word vector model, mapping each word to a high-dimensional vector; collecting a security domain vocabulary containing different security classification labels; for each security category word, taking the average value in the word vector space to obtain the word vector representative of that category; for new words, calculating the cosine similarity between the vector and the representative vectors of each category, using the following formula:

[0010]

[0011] Where A is the word vector of the vocabulary, A∈R n， Let B be an n-dimensional vector; let B be the representative word vector of security category i, where B∈R n， Let A be an n-dimensional vector; A·B be the dot product of A and B; ||A|| be the L2 norm of word vector A; ||B|| be the L2 norm of vector B; and the cosine of the angle between A and B, ranging from -1 to 1. The larger the cosine value, the closer the two vectors are in direction, and the higher the similarity. A similarity threshold is set, and words with a similarity greater than the threshold are considered to be highly related to the corresponding security category. For words with a similarity lower than the threshold, their security relevance is manually reviewed and judged. The results of the manual judgment are fed back, and the representative vectors of the relevant categories are adjusted. The above calculation, verification, and adjustment process is repeated until the security relevance calculation is accurate.

[0012] As a preferred embodiment of the project data security information hierarchical processing method described in this invention, the similarity threshold setting process is as follows: Collect verification word samples, and the security categories of these words are known; for each verification word, calculate the cosine similarity with the representative vector of each category; sort the cosine similarities and select the top K most similar categories; compare the matching situation of the top K categories with the actual categories of the words; adjust the similarity threshold, aiming to maximize the accuracy of the top K matches; construct the ROC curve of the verification samples, with the horizontal axis representing the false positive rate (FPR) and the vertical axis representing the true positive rate (TPR); divide the points in the curve where the TPR is higher than the average into reserve point set 1, and divide the points in the curve where the FPR is lower than the average into reserve point set 2; select the point with the largest difference between reserve point sets 1 and 2 on the same vertical line as the threshold point.

[0013] As a preferred embodiment of the project data security information hierarchical processing method described in this invention, the steps for determining the degree of security association are as follows: Constructing a security association knowledge graph; searching the graph for statement entities and returning all matching entity nodes; filtering out irrelevant meanings for ambiguous entities using context; marking the matching entity nodes in the statement; performing a depth-first search of the graph starting from the marked nodes; traversing node edges, inferring multi-hop association paths, and setting a maximum number of hops to avoid infinite traversal loops; returning all multi-hop association paths connecting statement entities; calculating the credibility weights of relations in the paths, accumulating the relation weights in the paths, and outputting the credibility of each association path; training a neural network to judge the credibility of the paths, where the credibility of path i is r. i， The path length is l i The input is [r] i ,l i The network is trained using labeled path data, where the labels represent the degree of security association of paths, y∈[0,1]. The optimized loss function is:

[0014]

[0015] in, For the predicted output; make the predicted output To approximate the actual label y, for a new path, input its [r] i ,l i [To the network, obtain the predicted security association probability] Summarize all paths Obtain the degree of security association of the statements.

[0016] As a preferred embodiment of the project data security information hierarchical processing method described in this invention, the process of calculating the relevance between data entities and the security terminology is as follows: comprehensively considering the matching degree of data format, semantic features, and security word roots; and matching a predefined security level based on the comprehensive judgment result. The process of comprehensively considering the matching degree of data format, semantic features, and security word roots is as follows: Checking the data format: if it is structured data, extracting field names for judgment; if it is unstructured data, performing text segmentation to extract keywords; performing semantic analysis on the extracted field names or keywords to determine whether they conform to name composition specifications; calculating... Calculate the semantic similarity between these keywords and predefined security roots, and determine the security level of the field name or keyword based on the semantic similarity. Sample the field content for further evaluation: use NLP technology to extract keywords from the sample content; calculate the term frequency (TF) of the extracted keywords in the sample text; perform semantic analysis on the sample content to determine if it contains sensitive information. If not, maintain the original security level; if the threshold for a small number of sensitive keywords is less than or equal to the threshold for a large number of sensitive keywords, increase the security level by one; if the threshold for contained sensitive information is equal to or greater than the threshold for a large number of sensitive keywords, increase the security level by two.

[0017] As a preferred embodiment of the project data security information classification processing method described in this invention, the process of matching predefined security levels is as follows: Data format is judged: Structured data: security coefficient S1 = 2; Semi-structured data: security coefficient S1 = 1; Unstructured data: security coefficient S1 = 0; Semantic similarity of field names or keywords is judged: Low similarity, security coefficient S2 = 1; Medium similarity, security coefficient S2 = 2; High similarity, security coefficient S2 = 3; The formula for calculating the security coefficient S is:

[0018] S = S1 + S2

[0019] The frequency of sensitive words in the sample content is judged as follows: when the frequency of sensitive words is less than the threshold for a small number of sensitive words, the security factor S remains unchanged; when the threshold for a small number of sensitive words is less than the threshold for a large number of sensitive words, the security factor S+1 is applied; when the frequency of sensitive words is greater than or equal to the threshold for a large number of sensitive words, the security factor S+2 is applied. The corresponding security level is matched according to the security factor S. If there is a conflict in the security factor, the priority is sample content > semantic similarity > data format.

[0020] Secondly, to further address the problems existing in data security classification, the present invention provides a project data security information classification processing system, which includes: a data acquisition module for acquiring relevant text data from a data compliance database and extracting security keywords; a security terminology construction module for processing the acquired security keywords and constructing a security terminology; verifying security relevance, adjusting the word vector space, and optimizing the terminology; and a data judgment module for performing format and semantic analysis on the data source, extracting features, matching the security terminology, calculating the degree of security relevance, applying the trained judgment model, evaluating the sensitivity of the data, and finally classifying the security level of the data.

[0021] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the project data security information classification processing method as described in the first aspect of the present invention.

[0022] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the project data security information classification processing method as described in the first aspect of the present invention.

[0023] The beneficial effects of this invention are as follows: it improves the intelligence level of data security classification, reduces manual workload, and enhances data processing efficiency; it applies natural language processing and deep learning technologies to improve the understanding and processing capabilities of unstructured data; by constructing a secure word vector space, it can more accurately determine the security correlation between words; multi-hop association reasoning based on knowledge graphs can reveal implicit data security correlations; it comprehensively evaluates multiple features, improving the accuracy and completeness of security classification; it dynamically calculates similarity thresholds and sensitive word frequency thresholds, avoiding the subjectivity of manual specification; and it provides a reliable basis for security classification, helping enterprises to better comply with data compliance requirements. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0025] Figure 1 This is a scenario flowchart of the project data security information classification processing method in Example 1. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0028] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0029] Example 1

[0030] Reference Figure 1 This is the first embodiment of the present invention, which provides a method for hierarchical processing of project data security information, including the following steps:

[0031] S1: Use the data compliance library to extract data security judgment rules and construct a data security thesaurus.

[0032] Preferably, relevant industry data compliance databases are collected, and text analysis of the compliance database is used to extract data security and risk control keywords from compliance regulations. The keywords are then processed to extract word roots, merge synonyms, and construct a data security thesaurus. The thesaurus contains a summary of keywords for data security judgment, such as confidentiality, password, and privacy. A security word vector space is constructed, the security relevance of words is calculated, and low-relevance words are filtered out.

[0033] Furthermore, constructing the security word vector space includes: using Word2Vec technology to train a word vector model, mapping each word to a high-dimensional vector; collecting a security domain vocabulary containing different security classification labels; for each security category word, taking the average in the word vector space to obtain the word vector representative of that category; for new words, calculating the cosine similarity between the vector and the representative vectors of each category, using the following formula:

[0034]

[0035] Where A is the word vector of the vocabulary, A∈R n， Let B be an n-dimensional vector; let B be the representative word vector of security category i, where B∈R n，Let A be an n-dimensional vector; A·B be the dot product of A and B; ||A|| be the L2 norm of word vector A; ||B|| be the L2 norm of vector B; and the cosine of the angle between A and B, ranging from -1 to 1. The larger the cosine value, the closer the two vectors are in direction, and the higher the similarity. A similarity threshold is set, and words with a similarity greater than the threshold are considered to be highly related to the corresponding security category. For words with low relevance, their security association is manually reviewed. The results of the manual review are fed back, and the representative vectors of the relevant categories are adjusted. The above calculation, manual verification, and adjustment process is repeated until the security association calculation is accurate. Finally, a word vector space model is formed, which can calculate the security classification relevance of new words and is used for lexicon construction.

[0036] Preferably, the threshold setting process is as follows: collect verification word samples, and the safety category of these words is known; for each verification word, calculate the cosine similarity with the representative vector of each category; sort the cosine similarity and select the top K most similar categories; compare the matching situation of the top K categories with the actual category of the word; adjust the similarity threshold, with the goal of maximizing the accuracy of the top K matches; construct the ROC curve of the verification samples, with the horizontal axis representing the false positive rate (FPR) and the vertical axis representing the true positive rate (TPR); divide the points in the curve where the TPR is higher than the average into reserve point set 1, divide the points in the curve where the FPR is lower than the average into reserve point set 2, and select the point with the largest difference between reserve point sets 1 and 2 on the same vertical line as the threshold point, that is, the similarity threshold point should make both TPR high and FPR low.

[0037] S2: Using natural language processing technology, based on a data security lexicon and semantic context analysis, an automatic hierarchical judgment model is established.

[0038] Collect a large amount of sample data and label it with security levels; use NLP technology to extract semantic features from the sample data; train a classification model, input the semantic features, and output the security level; continuously optimize the model to improve the accuracy of judgment.

[0039] Preferably, training a classification model includes the following steps: selecting a model suitable for the text classification task; inputting semantic features and outputting the security level probability of data samples; evaluating model performance using methods such as cross-validation; and selecting the best-performing model as the classification model.

[0040] Specifically, a semantic feature representation model BERT is pre-trained; the pre-trained semantic feature representation is used as the model input; the output layer is set as a fully connected layer with the number of nodes equal to the number of security levels; each node represents the predicted probability of the corresponding security level; a Softmax activation function is added to ensure that the sum of probabilities is 1; and the node weights are adjusted based on business needs for probability correction; for each security level, the cross-entropy loss is calculated; the loss of each level is averaged to obtain the multi-class cross-entropy; the loss weights are balanced by sampling by class; parameters are updated through backpropagation to minimize the loss; and the validation set loss is monitored during training; training is stopped early when the loss no longer decreases; regularization is used to prevent the model from becoming overly complex.

[0041] S3: Perform format and semantic recognition on the database data, and match security roots based on the recognition results to determine the degree of security association.

[0042] Preferably, NLP technology is used to extract the semantic content of the data fields.

[0043] Preferably, the process of matching security roots and determining the degree of security association includes: constructing a security association knowledge graph; parsing security semantics; performing security association reasoning; and verifying the association results of the reasoning.

[0044] Furthermore, the secure association reasoning process is as follows: Locating the statement entity: For the statement entity, search the graph and return all matching entity nodes; for ambiguous entities, filter out irrelevant meanings using context; mark the matched entity nodes in the statement; Reasoning the association path: Starting from the marked nodes, perform a depth-first search of the graph; traverse the node edges to reason out multi-hop association paths; set a maximum number of hops to avoid infinite traversal loops; and return all multi-hop association paths connecting the statement entities; Calculating path credibility: Calculate the credibility weight of the relations in the path and accumulate the relation weights in the path; output the credibility of each association path; Association judgment: Train the neural network to judge the path credibility; the network is input with features such as path credibility and length, let the credibility of path i be r. i， The path length is l i The input is [r] i ,l i The network is trained using labeled path data, where the labels represent the degree of security association of paths, y∈[0,1]. The loss function is optimized as follows:

[0045]

[0046] Make the predicted output Approximate the actual label y; for the new path, input its [r i ,l i [To the network, obtain the predicted security association probability] Summarize all paths Obtain the degree of security association of the statements.

[0047] S4: Calculate the relevance of data entities to the security thesaurus, automatically identify the sensitivity, and classify the security level of data entities based on their security characteristics.

[0048] S4.1: The degree of matching between comprehensive data format, semantic features and safe word roots.

[0049] Furthermore, the data format is checked: if it is structured data (database, table, etc.), the field names are extracted for judgment; if it is unstructured data (document, log, etc.), text segmentation is performed to extract keywords; semantic analysis is performed on the extracted field names / keywords: it is determined whether they conform to the name composition specifications (including keywords such as business concepts and data types); the semantic similarity between them and predefined security roots is calculated;

[0050] Based on semantic similarity, determine the security level of field names / keywords; sample the field content for further judgment: first, set a sensitive word threshold: use NLP technology to extract keywords from the sample content; calculate the term frequency (TF) of the extracted keywords in the sample text; fit a normal distribution to the term frequencies of all extracted keywords in the sample content, and determine the parameters μ and σ of the normal distribution. The threshold for a small number of sensitive keywords = μ + 0.5 × σ; the threshold for a large number of sensitive keywords = μ + σ; perform semantic analysis on the sample content to determine whether it contains sensitive information. If it does not contain sensitive information, maintain the original security level; if the threshold for a small number of sensitive keywords ≤ the threshold for contained sensitive information < the threshold for a large number of sensitive keywords, increase the security level by one; if the threshold for contained sensitive information ≥ the threshold for a large number of sensitive keywords, increase the security level by two.

[0051] The semantic similarity calculation process is as follows: A vector space model is constructed using word vector technology for security root words and field keywords; the semantic similarity between each field keyword and each security root word is calculated, and the maximum value is taken as the highest similarity between the field keyword and the security thesaurus; a normal distribution is fitted to the highest similarity of all fields to determine the parameters μ and σ of the normal distribution; the normal distribution is divided according to the three σ principle: similarity less than (μ-σ) is classified as low region; similarity between (μ-σ) and (μ+σ) is classified as medium region; and similarity greater than (μ+σ) is classified as high region.

[0052] S4.2: Based on the comprehensive judgment results, match the predefined security level.

[0053] Specifically, the data format is assessed as follows: Structured data: Security factor S1 = 2; Semi-structured data: Security factor S1 = 1; Unstructured data: Security factor S1 = 0; The semantic similarity of field names or keywords is assessed as follows: Low similarity: Security factor S2 = 1; Medium similarity: Security factor S2 = 2; High similarity: Security factor S2 = 3; The formula for calculating the security factor S is:

[0054] S = S1 + S2

[0055] The frequency of sensitive words in the sample content is judged as follows: when the frequency of sensitive words is less than the threshold for a small number of sensitive words, the security factor S remains unchanged; when the threshold for a small number of sensitive words is less than the threshold for a large number of sensitive words, the security factor S+1 is applied; when the frequency of sensitive words is greater than or equal to the threshold for a large number of sensitive words, the security factor S+2 is applied. The corresponding security level is matched according to the security factor S. If there is a conflict in the security factor, the priority is sample content > semantic similarity > data format.

[0056] S4.3: Output the security level classification for each data field.

[0057] This embodiment also provides a project data security information classification processing system, including a data acquisition module for collecting relevant text data from a data compliance database and extracting security keywords; a security terminology construction module for processing the collected security keywords and constructing a security terminology; verifying security relevance, adjusting the word vector space, and optimizing the terminology; and a data judgment module for performing format and semantic analysis on the data source, extracting features, matching the security terminology, calculating the degree of security relevance, applying the trained judgment model, assessing the sensitivity of the data, and finally classifying the security level of the data.

[0058] This embodiment also provides a computer device applicable to the project data security information classification processing method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the project data security information classification processing method proposed in the above embodiment.

[0059] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0060] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the data security information hierarchical processing method for the project as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0061] In summary, this invention improves the intelligence level of data security classification, reduces manual workload, and enhances data processing efficiency; it applies natural language processing and deep learning technologies to improve the understanding and processing capabilities of unstructured data; by constructing a secure word vector space, it can more accurately determine the security correlation between words; multi-hop association reasoning based on knowledge graphs can reveal implicit data security correlations; by comprehensively evaluating multiple features, it improves the accuracy and completeness of security classification; and by dynamically calculating similarity thresholds and sensitive word frequency thresholds, it avoids the subjectivity of manual specification. This invention provides a reliable basis for security classification, helping enterprises to better comply with data compliance requirements.

[0062] Example 2

[0063] Referring to Table 1, which is the second embodiment of the present invention, based on the first embodiment, experimental comparison data between the data security information classification processing method of the present invention and the prior art are provided to verify its beneficial effects.

[0064] First, a secure word vector space is constructed in this invention, as detailed in the table below:

[0065] Table 1. Security Word Vector Space

[0066] Name (0.1,0.3,0.5,...) Personal privacy ID number (0.2,0.1,0.4,...) Personal privacy Medical records (0.3,0.2,0.1,...) Sensitive information bill (0.4,0.2,0.3,...) General information

[0067] Word vectors are trained on vocabulary using word vector technology; the vocabulary and corresponding security categories are collected as follows:

[0068] Table 2. Correspondence between vocabulary and security categories

[0069] Personal privacy (0.15,0.2,0.45,...) Sensitive information (0.25,0.15,0.2,...) General information (0.3,0.25,0.35,...)

[0070] The method of this invention is used to calculate the representative word vector for each category. For example, for the new word "medical record", the word vector is (0.2, 0.25, 0.3, ...). The cosine similarity between its word vector and the representative vector of each category is calculated: personal privacy: 0.82; sensitive information: 0.88; general information: 0.76.

[0071] Furthermore, the safety classification of various terms in this invention is compared with that in the prior art as follows:

[0072] Table 3 Comparison of Classifications with Existing Technologies

[0073] Name (0.1,0.3,0.5,...) Personal privacy Personal privacy ID card number (0.2,0.1,0.4,...) Personal privacy Personal privacy phone number (0.3,0.2,0.2,...) General information Personal privacy Home address (0.4,0.3,0.1,...) General information Personal privacy Billing Information (0.2,0.4,0.3,...) Sensitive information General information Medical records (0.3,0.2,0.1,...) General information Sensitive information Outpatient records (0.25,0.15,0.2,...) General information Sensitive information Fee details (0.1,0.2,0.5,...) General information General information

[0074] As can be seen, by calculating the cosine similarity of word vectors, this invention can correctly classify words such as mobile phone numbers and home addresses that are easily misclassified by existing technologies, and can also more accurately determine the security category of words such as medical records than existing technologies. Compared with existing technologies that rely on subjective experience for security classification, this invention achieves more intelligent and accurate security classification through word vector similarity calculation. This technology can continuously optimize the word vector space and adapt to the judgment of newly added words, and has better scalability.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for hierarchical processing of project data security information, characterized in that: include: Data security judgment rules are extracted from the data compliance library to form a data security thesaurus; Using natural language processing technology, an automatic classification judgment model is established based on a data security lexicon and semantic context analysis; The database data is subjected to format and semantic recognition, and security roots are matched based on the recognition results to determine the degree of security association. Calculate the relevance of data entities to the security thesaurus, automatically identify the sensitivity, and classify the security level of data entities based on their security characteristics; The process of constructing the data security thesaurus includes the following steps: Collect relevant industry data compliance databases, use text analysis to extract data security and risk control keywords from compliance regulations; The keywords are processed, word roots are extracted, synonyms are merged, and a data security thesaurus is constructed. The thesaurus contains a summary of keywords for data security assessment; Constructing a secure word vector space, calculating the secure relevance of words, and filtering out low-relevance words; the construction of the secure word vector space includes: Using Word2Vec technology, a word vector model is trained to map each word to a high-dimensional vector; Collect a glossary of security terms, including different security category tags; For each security category word, the average value in the word vector space is taken to obtain the word vector representative of that category; For new words, calculate the cosine similarity between the vector and the representative vector of each category, using the following formula: ； Where A is the word vector of the vocabulary, A∈R n， Let B be an n-dimensional vector; let B be the representative word vector of security category i, where B∈R n， Let A be an n-dimensional vector; A·B is the dot product of A and B; ||A|| is the L2 norm of word vector A; ||B|| is the L2 norm of vector B. The cosine of the angle between A and B, with a range of [-1, 1]; A similarity threshold is set; if the similarity exceeds the threshold, the word is considered to be highly related to the corresponding security category. For words with a relevance score below the similarity threshold, a manual review is conducted to determine their safety relevance. Feedback on the human judgment results will be provided, and the representative vectors for the relevant categories will be adjusted accordingly. Repeat the above calculation, verification, and adjustment process until the security association calculation is accurate; The process of setting the similarity threshold is as follows: Collect and verify vocabulary samples, and the security categories of these words are known; For each validation word, calculate the cosine similarity with the representative vector of each category; Sort by cosine similarity and select the top K most similar categories; Compare the matching results of the top K categories with the actual categories of the words; Adjust the similarity threshold to maximize the accuracy of the top K matches; Construct the ROC curve for the validation samples, with the horizontal axis representing the false positive rate (FPR) and the vertical axis representing the true positive rate (TPR). Points in the curve with a TPR higher than the average are designated as preliminary point set 1, and points in the curve with an FPR lower than the average are designated as preliminary point set 2. The point containing the largest difference between the two sets of preparatory points 1 and 2 on the same vertical line is set as the threshold point.

2. The project data security information hierarchical processing method as described in claim 1, characterized in that: The steps for determining the degree of security association are as follows: Construct a security-related knowledge graph; For each statement entity, search the graph and return all matching entity nodes; For ambiguous entities, use context to filter out irrelevant meanings; The entity nodes matched in the tag statement; Starting from the marked nodes, perform a depth-first search of the graph; Traverse the edges of nodes, deduce multi-hop related paths, and set the maximum number of hops to avoid infinite traversal loops; Returns all multi-hop association paths of the join statement entity; Calculate the credibility weight of the relationship in the path, sum the relationship weights in the path, and output the credibility of each associated path; The neural network is trained to determine the credibility of a path. Let the credibility of path i be r. i， The path length is l i The input is [r] i ,l i The network is trained using labeled path data, where the labels represent the degree of security association of paths, y∈[0,1]. The optimized loss function is: L(y, ŷ) = (y - ŷ) 2 Where ŷ is the predicted output; to make the predicted output ŷ approximate the actual label y, for a new path, input its [r i , l i [Go to the network, obtain the predicted security association probability ŷ, summarize ŷ of all paths, and obtain the security association degree of the statement.] 3. The project data security information hierarchical processing method as described in claim 2, characterized in that: The process of calculating the relevance between the data entities and the security thesaurus is as follows: Based on the comprehensive judgment results, the system matches a predefined security level by considering the degree of matching between the data format, semantic features, and security root words. The process of determining the matching degree between the comprehensive data format, semantic features, and secure word roots is as follows: Check the data format: if it is structured data, extract the field names for judgment; if it is unstructured data, perform text segmentation to extract keywords. Perform semantic analysis on the extracted field names or keywords to determine whether they conform to the naming conventions; Calculate their semantic similarity to predefined security roots, and determine the security level of field names or keywords based on semantic similarity; Samples are extracted from the field content for further evaluation: NLP technology is used to extract keywords from the sample content; the term frequency (TF) of the extracted keywords in the sample text is calculated; and a normal distribution is fitted to the term frequencies of all extracted keywords in the sample content to determine the parameters μ and σ of the normal distribution. The threshold for a small number of sensitive keywords = μ + 0.5 × σ, and the threshold for a large number of sensitive keywords = μ + σ; Perform semantic analysis on the sample content to determine whether it contains sensitive information. If it does not, maintain the original security level. If the threshold for a small number of sensitive keywords is less than or equal to the threshold for a large number of sensitive keywords, the security level is increased by one level; if the threshold for a large number of sensitive keywords is greater than or equal to the threshold for a large number of sensitive keywords, the security level is increased by two levels.

4. The project data security information hierarchical processing method as described in claim 3, characterized in that: The process of matching the predefined security level is as follows: Determine the data format: Structured data: security factor S1 = 2; Semi-structured data: security factor S1 = 1; Unstructured data: security factor S1 = 0; The semantic similarity of field names or keywords is judged: when the similarity is low, the safety factor S2=1; when the similarity is medium, the safety factor S2=2; when the similarity is high, the safety factor S2=3. The formula for calculating the safety factor S is: S = S1 + S2 The frequency of sensitive words in the sample content is judged: when the frequency of sensitive words is less than the threshold for a small number of sensitive words, the safety factor S remains unchanged; When the frequency threshold of a small number of sensitive words is less than or equal to the frequency of a large number of sensitive words, the safety factor is S+1. When the frequency of sensitive words is greater than or equal to the threshold of a large number of sensitive words, the safety factor is S+2. The corresponding security level is matched based on the security factor S; if there is a conflict in the security factors, the priority is sample content > semantic similarity > data format.

5. A project data security information hierarchical processing system, based on the project data security information hierarchical processing method according to any one of claims 1 to 4, characterized in that: include, The data acquisition module is used to collect relevant text data from the data compliance database and extract security keywords; The security terminology building module is used to process the collected security keywords, build a security terminology, verify security relevance, adjust the word vector space, and optimize the terminology. The data judgment module is used to perform format and semantic analysis on the data source, extract features, match security terms to calculate the degree of security association, apply the trained judgment model, assess the sensitivity of the data, and finally classify the security level of the data.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the project data security information hierarchical processing method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the project data security information hierarchical processing method according to any one of claims 1 to 4.

Images