Data classification and grading method and system based on artificial intelligence

Abstract

The invention discloses a data classification and grading method and system based on artificial intelligence, and relates to the technical field of data processing, and the method comprises the steps: obtaining multi-source data, and carrying out the cleaning and standardization processing; constructing a multi-modal data feature extraction model, extracting data multi-source features through a deep learning network, and generating a unified 512-dimensional global feature vector; determining a feature weight by combining a dynamic analytic hierarchy process and fuzzy comprehensive evaluation, calculating a comprehensive sensitivity feature score, a comprehensive service feature score and a comprehensive safety risk feature score, simultaneously outputting a comprehensive score, and performing classification and grading according to the comprehensive score; setting a sensitivity priority rule to directly divide high-sensitivity data; generating a hierarchical security policy; triggering model updating through dynamic monitoring; according to the method, the problems that a traditional classification and grading method depends on static rules and lacks multi-dimensional dynamic evaluation are solved, and the accuracy and strategy adaptability of data classification and grading are improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a data classification and grading method and system based on artificial intelligence. Background Technology

[0002] As enterprises undergo digital transformation, the scale and complexity of data surge, revealing the following shortcomings of traditional rule-based or manually defined classification methods:

[0003] Relying on preset keywords or fixed thresholds makes it impossible to adapt to business changes and new threats; it only focuses on sensitivity or business attributes, ignoring the correlation between security risks and data characteristics; the model lacks a dynamic update mechanism, making it difficult to respond to new data types or risk changes, and the homogenization of hierarchical strategies fails to reflect the differences in storage, access, and transmission requirements of different levels of data.

[0004] While existing technologies have achieved data classification and grading to a certain extent, they suffer from problems such as insufficient multimodal feature fusion and weight calculations that are detached from actual business scenarios.

[0005] Therefore, there is an urgent need for an intelligent classification and grading method that integrates multi-dimensional dynamic evaluation and adaptive updating. Summary of the Invention

[0006] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a data classification and grading method and system based on artificial intelligence, to solve the following technical problem:

[0007] Relying on preset keywords or fixed thresholds makes it impossible to adapt to business changes and new threats; it only focuses on sensitivity or business attributes, ignoring the correlation between security risks and data characteristics; the model lacks a dynamic update mechanism, making it difficult to respond to new data types or risk changes, and the homogenization of hierarchical strategies fails to reflect the differences in storage, access, and transmission requirements of different levels of data.

[0008] To address the above problems, a first aspect of the present invention provides a data classification and grading method based on artificial intelligence, comprising the following steps:

[0009] S1: Acquire multi-source data within the enterprise and perform data cleaning and standardization processing;

[0010] S2: Construct a multimodal data feature extraction model, extract multi-source features of data through a deep learning network, and generate a unified 512-dimensional global feature vector;

[0011] S3: Combining dynamic hierarchical analysis with fuzzy comprehensive evaluation, calculate the dynamic weights of sensitivity, business needs and security risks, and calculate the comprehensive sensitivity feature score, comprehensive business feature score and comprehensive security risk feature score. Output a comprehensive score and classify the data level and category according to the comprehensive score.

[0012] S4: Set a sensitivity priority rule. When the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level is directly classified as high-sensitivity data, corresponding to level 4.

[0013] S5: Generate a hierarchical security policy based on the data classification and grading results. The policy includes differentiated configurations of data storage encryption strength, access permission granularity, and transmission encryption protocol.

[0014] S6: Trigger model updates through dynamic monitoring mechanism: Calculate the business demand change coefficient and the security risk change coefficient. If the sum of the two coefficients is greater than a preset threshold, start model retraining. The retraining adopts an incremental learning framework.

[0015] Preferably, the data cleaning and standardization process in step S1 includes the following steps:

[0016] Data cleaning and standardization processes employ a multi-strategy fusion mechanism: for structured data, Z-Score standardization maps numerical fields to the [0, 1] interval, while one-hot encoding combined with frequency weighting is used for categorical fields;

[0017] For unstructured data, a semantic vector transformation based on the BERT model is used, and duplicate text is filtered by cosine similarity, with the duplicate detection threshold set to 0.92.

[0018] Histogram equalization was used for preprocessing of image data, and Mel-spectral feature transformation was used for audio data.

[0019] Preferably, the multimodal data feature extraction model in step S2 includes the following steps:

[0020] The multimodal data feature extraction model adopts a cross-membrane attention fusion architecture;

[0021] Text data is processed using a bidirectional LSTM to extract temporal features with a hidden dimension of 256; image data is processed using the conv5 layer of ResNet50 to extract visual features with an output dimension of 2048; and structured data is processed using three fully connected layers to extract statistical features.

[0022] The three types of features generate a global feature vector through a self-attention mechanism, with the feature dimension being uniformly 512 and satisfying the L2 norm constraint.

[0023] Preferably, the dynamic hierarchical analysis method in step S3 includes the following steps:

[0024] The judgment matrix update cycle is synchronized with the business iteration cycle, and a time decay function is introduced to correct the weights of the historical judgment matrix.

[0025] A three-level evaluation set {low, medium, high} is adopted, and the corresponding membership functions adopt descending semi-normal, normal, and ascending semi-normal distributions, respectively. The standard deviation of the membership function for the sensitive dimension is set to 0.15, that for the business dimension is set to 0.20, and that for the security risk dimension is set to 0.18.

[0026] Preferably, the comprehensive sensitivity feature score includes the following steps:

[0027] By constructing a well-developed multimodal data feature extraction model, semantic features of personal information, business passwords, and regulatory-sensitive information are extracted from the data.

[0028] An algorithm based on semantic similarity is used to match the extracted sensitive entities with a preset sensitive word library, and a basic score is assigned according to the matching degree.

[0029] By combining the frequency of occurrence and distribution density of sensitive entities, a comprehensive sensitivity feature score is obtained, specifically:

[0030] S = S base +α1·f+α2·d

[0031] Where S is the comprehensive sensitivity feature score, S base The base score is f, the frequency of occurrence is d, the distribution density is d, and α1 and α2 are the corresponding weights. The distribution density is obtained by dividing the number of sensitive entities by the total length of the data.

[0032] Preferably, the comprehensive business feature score includes the following steps:

[0033] Based on the enterprise business architecture map, identify the participation and influence of data in core business processes;

[0034] By analyzing the correlation between business process nodes, the Pearson correlation coefficient between the data and key business indicators is calculated, and the absolute value of the correlation coefficient is mapped to a basic business score of 0-1.

[0035] By determining the irreplaceability of the data and using redundancy to correct the basic business score, a comprehensive business feature score is obtained, specifically:

[0036] B = B base *100*(1-r)

[0037] Where B is the comprehensive business characteristic score, B baseThe score is based on basic business operations, and r represents data redundancy.

[0038] Preferably, the comprehensive security risk characteristic score includes the following steps:

[0039] The threat matrix analysis method is used to calculate the internal and external threat risks faced by the data.

[0040] External risk dimensions include the price of data breaches in the black market and the probability of similar companies experiencing data security incidents; internal risk dimensions include the frequency of unauthorized access and the level of vulnerability in data transmission encryption strength.

[0041] Each safety risk indicator is quantified into a sub-score of 0-1 points, and then a comprehensive safety risk characteristic score is obtained by weighting and summing the scores based on the probability of risk occurrence and the degree of impact. Specifically:

[0042]

[0043] Where R is the comprehensive security risk characteristic score, m is the total number of security risk indicators, and p j Let e ​​be the probability of occurrence of the j-th type of risk. j β represents the impact level of the j-th type of risk. j is the weight of the j-th type of risk.

[0044] Preferably, the comprehensive score in step S3 includes the following steps:

[0045] The overall score is as follows:

[0046] Score=ω S ·S+ω B ·B+ω R ·R

[0047] Where Score is the overall score, S is the overall sensitivity feature score, B is the overall business feature score, R is the overall security risk feature score, and ω is the overall security risk feature score. S ω B and ω R These are the weight coefficients corresponding to the scores of the three types of features, and the weight coefficients are dynamically generated by DAHP.

[0048] Data is classified into levels and categories based on the comprehensive score, including: if the comprehensive score is ≤30, the data is level 1, corresponding to public data; if 30 < comprehensive score ≤60, the data is level 2, corresponding to internal data; if 60 < comprehensive score ≤80, the data is level 3, corresponding to sensitive data; if the comprehensive score is >80, the data is level 4, corresponding to highly sensitive data.

[0049] If the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level will be directly classified as high-sensitivity data, corresponding to level 4.

[0050] Preferably, step S6 includes the following steps:

[0051] The coefficient for the change in business requirements is specifically as follows:

[0052]

[0053] Where c1 is the business demand variation coefficient, v current v is the current business feature vector. history This is a historical business feature vector;

[0054] The aforementioned safety risk change coefficient is specifically:

[0055]

[0056] Where c2 is the safety risk change coefficient, x current，i Let x be the value of the current i-th security risk indicator. base，i Let m be the baseline value of the i-th security risk indicator, and m be the total number of security risk indicators.

[0057] The adaptive threshold is specifically:

[0058] T = 0.5 + 0.3 * max(c1, c2)

[0059] Where T is the adaptive threshold, c1 is the business demand change coefficient, and c2 is the security risk change coefficient;

[0060] The retraining conditions for the model are: c1+c2>T and the proportion of new data ≥30%, wherein the new data must include data types added within the last 3 months.

[0061] A second aspect of the present invention provides an artificial intelligence-based data classification and grading system, comprising the following modules:

[0062] Multi-source data cleaning and structuring module: Acquires multi-source data from within the enterprise and performs data cleaning and standardization processing;

[0063] Cross-modal feature extraction and fusion module: Constructs a multimodal data feature extraction model, extracts multi-source features of data through a deep learning network, and generates a unified 512-dimensional global feature vector;

[0064] Dynamic weight calculation and comprehensive score module: Combining dynamic hierarchical analysis and fuzzy comprehensive evaluation, it calculates the dynamic weights of sensitivity, business requirements and security risks, and calculates the comprehensive sensitivity feature score, comprehensive business feature score and comprehensive security risk feature score, outputs the comprehensive score, and classifies the data level and category according to the comprehensive score;

[0065] Sensitivity Priority Rule Determination Module: Set sensitivity priority rules. When the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level is directly classified as high-sensitivity data, corresponding to level 4.

[0066] Hierarchical security policy generation module: Generates hierarchical security policies based on data classification and grading results. The policies include differentiated configurations of data storage encryption strength, access permission granularity, and transmission encryption protocols.

[0067] Dynamic monitoring and model update triggering module: Triggers model updates through dynamic monitoring mechanism: Calculates the business requirement change coefficient and the security risk change coefficient. If the sum of the two coefficients is greater than a preset threshold, the model retraining is started. The retraining adopts an incremental learning framework.

[0068] The beneficial effects of this invention are:

[0069] This invention solves the problem of heterogeneity in multi-source data by uniformly extracting features from text, images, and structured data through a cross-membrane attention model.

[0070] This invention combines DAHP with comprehensive fuzzy evaluation to adjust the weights of sensitivity, business, and security risks in real time, avoiding misjudgments caused by fixed rules. At the same time, it calculates the comprehensive sensitivity feature score, comprehensive business feature score, and comprehensive security risk feature score, and adds them together after weighting to obtain a comprehensive score. The data is classified and graded according to the comprehensive score. Furthermore, a fast classification channel is set up for highly sensitive data to ensure that key data is directly classified into the highest level.

[0071] This invention improves the granularity of control by generating hierarchical configurations of encryption strength, permission granularity, and transmission protocol based on data level, and ensures the model's responsiveness to new data types and threats by triggering incremental learning through business and security risk change coefficients. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0073] Figure 2 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Please see Figure 1 As shown, this invention is a data classification and grading method based on artificial intelligence, comprising the following steps:

[0076] S1: Acquire multi-source data within the enterprise and perform data cleaning and standardization processing;

[0077] S2: Construct a multimodal data feature extraction model, extract multi-source features of data through a deep learning network, and generate a unified 512-dimensional global feature vector;

[0078] S3: Combining dynamic hierarchical analysis with fuzzy comprehensive evaluation, calculate the dynamic weights of sensitivity, business needs and security risks, and calculate the comprehensive sensitivity feature score, comprehensive business feature score and comprehensive security risk feature score. Output a comprehensive score and classify the data level and category according to the comprehensive score.

[0079] S4: Set a sensitivity priority rule. When the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level is directly classified as high-sensitivity data, corresponding to level 4.

[0080] S5: Generate a hierarchical security policy based on the data classification and grading results. The policy includes differentiated configurations of data storage encryption strength, access permission granularity, and transmission encryption protocol.

[0081] S6: Trigger model updates through dynamic monitoring mechanism: Calculate the business demand change coefficient and the security risk change coefficient. If the sum of the two coefficients is greater than a preset threshold, start model retraining. The retraining adopts an incremental learning framework.

[0082] Specifically, the graded security policy in step S5 includes differentiated control measures, including: Level 1: Data is stored in transparent encryption and is accessible to all internal staff; Level 2: Data is subject to role-based access control and supports department-based access auditing; Level 3: Data requires two-factor authentication and operation logs are kept for no less than 180 days; and Level 4: Data is stored in a hardware encryption machine and access requires three levels of approval and is watermarked throughout the process.

[0083] In one embodiment of the present invention, the data cleaning and standardization process in step S1 includes the following steps:

[0084] Data cleaning and standardization processes employ a multi-strategy fusion mechanism: for structured data, Z-Score standardization maps numerical fields to the [0, 1] interval, while one-hot encoding combined with frequency weighting is used for categorical fields;

[0085] For unstructured data, a semantic vector transformation based on the BERT model is used, and duplicate text is filtered by cosine similarity, with the duplicate detection threshold set to 0.92.

[0086] Histogram equalization was used for preprocessing of image data, and Mel-spectral feature transformation was used for audio data.

[0087] Specifically, for structured data, the mean and standard deviation of the fields are calculated, and Z-score standardization is used to map the values ​​to the [0, 1] interval, truncating negative or outlier values. For categorical data, one-hot encoding is used to convert the category values ​​into binary vectors; for example, category A corresponds to [1, 0, 0] and category B corresponds to [0, 1, 0]. The frequency of occurrence of each category is counted, and the weight is calculated. Where, q i f represents the weight corresponding to the i-th category. i Let f be the frequency of occurrence of the i-th category, and max(f) i The maximum frequency of occurrence in this category is used to multiply the original one-hot encoded vector by the corresponding weight to obtain the weighted one-hot vector. For unstructured data, a pre-trained BERT model is used to extract 768-dimensional semantic vectors from text sentences. Long texts are segmented, and the mean of the vectors in each segment is taken as the global representation. The cosine similarity value of each pair of text vectors is calculated. If the similarity is ≥0.92, different thresholds (0.85, 0.88, 0.90, 0.92, and 0.95) are tested on labeled datasets. At 0.92, the recall rate is >90% and the false positive rate is <5%, which is the best performance. If the similarity is ≥0.92, it is marked as a duplicate and one of them is retained. For image data, color images are converted to grayscale images, grayscale histograms are calculated, the frequency of each pixel value is counted, and pixel values ​​are mapped using the histogram equalization formula. Specifically: Among them, S k The values ​​are the mapped pixel values, where L is the number of gray levels and n is the number of gray levels. j Let j be the frequency of pixel value j, round(·) be the rounding function, k be the gray level, and output an image with enhanced contrast. For audio data processing, the sampling rate is uniformly set to 16kHz, mono format, framed (e.g., 25ms frame length, 10ms frame shift) and windowed, 40 Mel filters are set to cover the frequency range, the energy of each Mel filter is calculated, the logarithm of the output of each Mel filter is taken, and a fixed-dimensional Mel spectral feature matrix is ​​output, for example, 40 dimensions per frame. Multi-strategy fusion is used to achieve unified standardization of multimodal data, providing high-quality input for subsequent feature extraction and modeling.

[0088] In one embodiment of the present invention, the multimodal data feature extraction model in step S2 includes the following steps:

[0089] The multimodal data feature extraction model adopts a cross-membrane attention fusion architecture;

[0090] Text data is processed using a bidirectional LSTM to extract temporal features with a hidden dimension of 256; image data is processed using the conv5 layer of ResNet50 to extract visual features with an output dimension of 2048; and structured data is processed using three fully connected layers to extract statistical features.

[0091] The three types of features generate a global feature vector through a self-attention mechanism, with the feature dimension being uniformly 512 and satisfying the L2 norm constraint.

[0092] Specifically, the text data is cleaned and segmented or tokenized to convert it into word vector representations. The preprocessed text sequence is then input into a bidirectional LSTM model. The hidden dimension of the LSTM is set to 256, meaning both the forward and backward LSTM outputs have a dimension of 256. The final output of the bidirectional LSTM is the concatenation of the hidden states from both directions, resulting in a feature vector with a dimension of 256 × 2 = 512. Pooling is then performed on the LSTM output to obtain a fixed-length text feature vector. For image data, standardization is performed. The image is input into a pre-trained ResNet50 model, using the ResNet50's conv5 layer as a feature extractor with an output dimension of 2048. Global average pooling is then performed on the output of the conv5 layer to obtain image features. The process involves standardizing structured data into fixed-length feature vectors. These vectors are then fed into a three-layer fully connected network with ReLU activation functions at 1024, 512, and 256 output dimensions. The resulting structured data feature vectors are mapped to a unified 512-dimensional matrix. This unified matrix is ​​then concatenated and applied to a self-attention mechanism: the query, key, and value are calculated, attention weights are computed, and a weighted sum is obtained to form a global feature vector. This vector is then average-pooled to achieve the final global feature vector. Finally, the global feature vector is L2-norm normalized to ensure an L2 norm of 1, resulting in a global feature vector that satisfies the L2 norm constraint.

[0093] In one embodiment of the present invention, the dynamic hierarchical analysis method in step S3 includes the following steps:

[0094] The judgment matrix update cycle is synchronized with the business iteration cycle, and a time decay function is introduced to correct the weights of the historical judgment matrix.

[0095] A three-level evaluation set {low, medium, high} is adopted, and the corresponding membership functions adopt descending semi-normal, normal, and ascending semi-normal distributions, respectively. The standard deviation of the membership function for the sensitive dimension is set to 0.15, that for the business dimension is set to 0.20, and that for the security risk dimension is set to 0.18.

[0096] Specifically, an evaluation indicator system is defined, including sensitivity, business, and security risk dimensions. An initial hierarchical structure is constructed, clarifying the hierarchical relationships between indicators. Based on historical data, an initial judgment matrix is ​​constructed to represent the relative importance of each indicator. The update cycle of the judgment matrix is ​​set to be consistent with the business iteration cycle, for example, updated monthly. The update process of the judgment matrix is ​​triggered with each business iteration. A time decay function is defined, specifically: f(t) = e -λt Where f(t) is the time decay function, t is the number of cycles from the current time, and λ is the decay rate parameter with a value of 0.1. The weights of the historical judgment matrix are adjusted as follows: in, The weights after time decay. Let f(t) be the initial weight and f(t) be the time decay function. Based on the latest business data, update the judgment matrix and normalize it to ensure the weight sum is 1. Calculate the updated weight vector using the eigenvector method, and multiply it with the time decay function to obtain the final dynamic weight. Use a three-level comment set {low, medium, high} to correspond to the three levels of fuzzy evaluation. Design membership functions using low: decreasing semi-normal distribution, medium: normal distribution, and high: increasing semi-normal distribution. Calculate the membership degree of each evaluation indicator belonging to "low," "medium," and "high," and combine the membership degrees of all indicators into a membership degree matrix. Combine the final dynamic weights of each indicator obtained through dynamic hierarchical analysis with the membership degree matrix for comprehensive evaluation. Output the comprehensive evaluation result, representing the overall evaluation's membership degree to "low," "medium," and "high." Based on the principle of maximum membership degree, select the comment level corresponding to the maximum comprehensive evaluation result as the final result.

[0097] In one embodiment of the present invention, the comprehensive sensitivity feature score includes the following steps:

[0098] By constructing a well-developed multimodal data feature extraction model, semantic features of personal information, business passwords, and regulatory-sensitive information are extracted from the data.

[0099] An algorithm based on semantic similarity is used to match the extracted sensitive entities with a preset sensitive word library, and a basic score is assigned according to the matching degree.

[0100] By combining the frequency of occurrence and distribution density of sensitive entities, a comprehensive sensitivity feature score is obtained, specifically:

[0101] S = S base +α1·f+α2·d

[0102] Where S is the comprehensive sensitivity feature score, S base The base score is f, the frequency of occurrence is d, the distribution density is d, and α1 and α2 are the corresponding weights. The distribution density is obtained by dividing the number of sensitive entities by the total length of the data.

[0103] Specifically, sensitive entities are identified from multimodal features, including extracting sensitive entities from text using named entity recognition models, identifying sensitive text in images using image OCR technology, and extracting sensitive information from audio using speech-to-text technology, outputting a list of sensitive entities E = {e1, e2, ..., e...}. n}; Construct a sensitive word database C = {c1, c2, ..., c m The list contains keywords related to personal information, business passwords, and regulatory sensitive information, such as C = ID number, password, confidential, privacy. For each sensitive entity, the semantic similarity between it and each word in the sensitive word list is calculated. Specifically, a pre-trained language model is used to calculate the semantic embedding vector between the sensitive entity and each word in the sensitive word list, the cosine similarity is calculated, the maximum similarity is selected as the matching degree of the sensitive entity, and the matching degree list M = {m1, m2, ..., m} is output. n A base score is assigned based on the matching degree, specifically: S base_i =m i *max_s, where S base_i Let m be the base score for the i-th sensitive entity. i Let max_s be the matching degree of the i-th sensitive entity, and 100 be the upper limit of the base score. Count the number of times each sensitive entity appears in the data, calculate the distribution density of the sensitive entity, and combine the base score, frequency of occurrence and distribution density to calculate the comprehensive sensitivity feature score, where α1 and α2 are the corresponding weights with values ​​of 0.6 and 0.4, respectively. Normalize the comprehensive sensitivity score to ensure that the score is within the range of 0-100.

[0104] In one embodiment of the present invention, the comprehensive business feature scoring includes the following steps:

[0105] Based on the enterprise business architecture map, identify the participation and influence of data in core business processes;

[0106] By analyzing the correlation between business process nodes, the Pearson correlation coefficient between the data and key business indicators is calculated, and the absolute value of the correlation coefficient is mapped to a basic business score of 0-1.

[0107] By determining the irreplaceability of the data and using redundancy to correct the basic business score, a comprehensive business feature score is obtained, specifically:

[0108] B = B base *100*(1-r)

[0109] Where B is the comprehensive business characteristic score, B base The score is based on basic business operations, and r represents data redundancy.

[0110] Specifically, business process modeling tools are used to draw core business process diagrams for the enterprise, labeling the input, output, and processing logic of each business process node, identifying the role of data in the business process (e.g., input data, intermediate results, and output data), and outputting a business architecture diagram. The frequency of data occurrence in business process nodes is statistically analyzed to assess the data's dependence on these nodes. High dependence means the data is a necessary input or output of the business process node; low dependence means the data is an optional input or intermediate result. A data participation score is obtained by multiplying the frequency of occurrence by the dependence. Key business indicators (KPIs) are extracted from the business architecture diagram, such as revenue, cost, and profit in financial indicators, and order processing time and customer satisfaction in operational indicators. A list of KPIs is output, historical values ​​of data and KPIs are collected, Pearson correlation coefficients are calculated, and the absolute value of the correlation coefficient is taken to output a list of correlation coefficients between data and each KPI. The absolute value of the correlation coefficient is mapped to a basic business score of 0-1. The number of times data is repeated in the business process is statistically analyzed to calculate redundancy. A comprehensive business characteristic score is calculated by combining the basic business score and the redundancy.

[0111] In one embodiment of the present invention, the comprehensive security risk feature score includes the following steps:

[0112] The threat matrix analysis method is used to calculate the internal and external threat risks faced by the data.

[0113] External risk dimensions include the price of data breaches in the black market and the probability of similar companies experiencing data security incidents; internal risk dimensions include the frequency of unauthorized access and the level of vulnerability in data transmission encryption strength.

[0114] Each safety risk indicator is quantified into a sub-score of 0-1 points, and then a comprehensive safety risk characteristic score is obtained by weighting and summing the scores based on the probability of risk occurrence and the degree of impact. Specifically:

[0115]

[0116] Where R is the comprehensive security risk characteristic score, m is the total number of security risk indicators, and p j Let e ​​be the probability of occurrence of the j-th type of security risk. jβ represents the impact level of the j-th type of security risk. j Let be the weight of the j-th type of security risk.

[0117] Specifically, each security risk dimension is converted into a quantifiable indicator, and normalized to a score of 0-1 using a min-max normalization formula. Specifically: Data leakage black market price: Based on market data, the price is mapped to a score of 0-1; the higher the price, the higher the score. Probability of similar company incidents: The frequency of similar company incidents is statistically analyzed and mapped to a score of 0-1; the higher the frequency, the higher the score. Unauthorized access frequency: Internal access logs are statistically analyzed and mapped to a score of 0-1; the higher the frequency, the higher the score. Encryption strength defect level: The severity of encryption technology defects is assessed and mapped to a score of 0-1; the more severe the defect, the higher the score. The quantified security risk indicator matrix is ​​output. Two dimensions of the threat matrix are defined: Security risk occurrence probability: Used to assess the likelihood of each type of risk occurring; Risk impact degree: Used to assess the impact of each type of risk after it occurs. As shown in Table 1:

[0118] Security Risk Categories Probability of security risks Degree of impact of security risks Data breach black market prices 0.8 0.9 Probability of similar events 0.7 0.7 Unauthorized access frequency 0.6 0.8 Encryption strength defects 0.5 0.6

[0119] For each type of security risk, a score is calculated, and weights are assigned to each type of risk according to business needs. The weights for data leakage black market prices, probability of similar enterprise incidents, frequency of unauthorized access, and encryption strength defects are 0.3, 0.2, 0.3, and 0.2, respectively. The scores of all security risk categories are weighted and summed to obtain the comprehensive security risk feature score.

[0120] In one embodiment of the present invention, the comprehensive score in step S3 includes the following steps:

[0121] The overall score is as follows:

[0122] Score=ω S ·S+ω B ·B+ω R ·R

[0123] Where Score is the overall score, S is the overall sensitivity feature score, B is the overall business feature score, R is the overall security risk feature score, and ω is the overall security risk feature score. S ω B and ω R These are the weight coefficients corresponding to the scores of the three types of features, and the weight coefficients are dynamically generated by DAHP.

[0124] Data is classified into levels and categories based on the comprehensive score, including: if the comprehensive score is ≤30, the data is level 1, corresponding to public data; if 30 < comprehensive score ≤60, the data is level 2, corresponding to internal data; if 60 < comprehensive score ≤80, the data is level 3, corresponding to sensitive data; if the comprehensive score is >80, the data is level 4, corresponding to highly sensitive data.

[0125] If the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level will be directly classified as high-sensitivity data, corresponding to level 4.

[0126] Specifically, the weighting coefficients of the three categories of comprehensive feature scores are dynamically generated using the analytic hierarchy process (AHP). The comprehensive score is obtained by weighted summation based on these weighting coefficients. A preset sensitivity threshold is defined to directly identify highly sensitive data. Specifically, through testing with historical data, the score distribution of highly sensitive data is statistically analyzed, and a threshold that can cover more than 95% of highly sensitive data is selected. For example, if experiments show that data with S > 0.7 are all highly sensitive data, then the threshold is set to 0.7.

[0127] In one embodiment of the present invention, step S6 includes the following steps:

[0128] The coefficient for the change in business requirements is specifically as follows:

[0129]

[0130] Where c1 is the business demand variation coefficient, v current v is the current business feature vector. history This is a historical business feature vector;

[0131] The aforementioned safety risk change coefficient is specifically:

[0132]

[0133] Where c2 is the safety risk change coefficient, x current，i Let x be the value of the current i-th security risk indicator. base，i Let m be the baseline value of the i-th security risk indicator, and m be the total number of security risk indicators.

[0134] The adaptive threshold is specifically:

[0135] T = 0.5 + 0.3 * max(c1, c2)

[0136] Where T is the adaptive threshold, c1 is the business demand change coefficient, and c2 is the security risk change coefficient;

[0137] The retraining conditions for the model are: c1+c2>T and the proportion of new data ≥30%, wherein the new data must include data types added within the last 3 months.

[0138] Specifically, real-time collection of business characteristic data and security risk characteristic data is performed. The business characteristic data includes the current business characteristic vector and the historical business characteristic vector, while the security risk characteristic data includes the value of the current i-th security risk indicator, the baseline value of the i-th security risk indicator, and the total number of security risk indicators. The monitoring cycle and data range are set to ensure that new data collection covers the types of data added in the past three months.

[0139] Please see Figure 2 As shown, this invention is a data classification and grading system based on artificial intelligence, comprising the following modules:

[0140] Multi-source data cleaning and structuring module: Acquires multi-source data from within the enterprise and performs data cleaning and standardization processing;

[0141] Cross-modal feature extraction and fusion module: Constructs a multimodal data feature extraction model, extracts multi-source features of data through a deep learning network, and generates a unified 512-dimensional global feature vector;

[0142] Dynamic weight calculation and comprehensive score module: Combining dynamic hierarchical analysis and fuzzy comprehensive evaluation, it calculates the dynamic weights of sensitivity, business requirements and security risks, and calculates the comprehensive sensitivity feature score, comprehensive business feature score and comprehensive security risk feature score, outputs the comprehensive score, and classifies the data level and category according to the comprehensive score;

[0143] Sensitivity Priority Rule Determination Module: Set sensitivity priority rules. When the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level is directly classified as high-sensitivity data, corresponding to level 4.

[0144] Hierarchical security policy generation module: Generates hierarchical security policies based on data classification and grading results. The policies include differentiated configurations of data storage encryption strength, access permission granularity, and transmission encryption protocols.

[0145] Dynamic monitoring and model update triggering module: Triggers model updates through dynamic monitoring mechanism: Calculates the business requirement change coefficient and the security risk change coefficient. If the sum of the two coefficients is greater than a preset threshold, the model retraining is started. The retraining adopts an incremental learning framework.

[0146] The above embodiments are only used to illustrate the technical methods 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 methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A data classification and grading method based on artificial intelligence, characterized in that, Includes the following steps: S1: Acquire multi-source data within the enterprise and perform data cleaning and standardization processing; S2: Construct a multimodal data feature extraction model, extract multi-source features of data through a deep learning network, and generate a unified 512-dimensional global feature vector; S3: Combining dynamic hierarchical analysis with fuzzy comprehensive evaluation, calculate the dynamic weights of sensitivity, business needs and security risks, and calculate the comprehensive sensitivity feature score, comprehensive business feature score and comprehensive security risk feature score. Output a comprehensive score and classify the data level and category according to the comprehensive score. S4: Set a sensitivity priority rule. When the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level is directly classified as high-sensitivity data, corresponding to level 4. S5: Generate a hierarchical security policy based on the data classification and grading results. The policy includes differentiated configurations of data storage encryption strength, access permission granularity, and transmission encryption protocol. S6: Trigger model updates through dynamic monitoring mechanism: Calculate the business demand change coefficient and the security risk change coefficient. If the sum of the two coefficients is greater than a preset threshold, start model retraining. The retraining adopts an incremental learning framework.

2. The data classification and grading method based on artificial intelligence according to claim 1, characterized in that, The data cleaning and standardization process in step S1 includes the following steps: Data cleaning and standardization processes employ a multi-strategy fusion mechanism: for structured data, Z-Score standardization maps numerical fields to the [0, 1] interval, while one-hot encoding combined with frequency weighting is used for categorical fields; For unstructured data, a semantic vector transformation based on the BERT model is used, and duplicate text is filtered by cosine similarity, with the duplicate detection threshold set to 0.

92. Histogram equalization was used for preprocessing of image data, and Mel-spectral feature transformation was used for audio data.

3. The data classification and grading method based on artificial intelligence according to claim 1, characterized in that, The multimodal data feature extraction model in step S2 includes the following steps: The multimodal data feature extraction model adopts a cross-membrane attention fusion architecture; Text data is processed using a bidirectional LSTM to extract temporal features with a hidden dimension of 256; image data is processed using the conv5 layer of ResNet50 to extract visual features with an output dimension of 2048; and structured data is processed using three fully connected layers to extract statistical features. The three types of features generate a global feature vector through a self-attention mechanism, with the feature dimension being uniformly 512 and satisfying the L2 norm constraint.

4. The data classification and grading method based on artificial intelligence according to claim 1, characterized in that, The dynamic hierarchical analysis method in step S3 includes the following steps: The judgment matrix update cycle is synchronized with the business iteration cycle, and a time decay function is introduced to correct the weights of the historical judgment matrix. A three-level evaluation set {low, medium, high} is adopted, and the corresponding membership functions adopt descending semi-normal, normal, and ascending semi-normal distributions, respectively. The standard deviation of the membership function for the sensitive dimension is set to 0.15, that for the business dimension is set to 0.20, and that for the security risk dimension is set to 0.

18.

5. The data classification and grading method based on artificial intelligence according to claim 1, characterized in that, The comprehensive sensitivity feature score includes the following steps: By constructing a well-developed multimodal data feature extraction model, semantic features of personal information, business passwords, and regulatory-sensitive information are extracted from the data. An algorithm based on semantic similarity is used to match the extracted sensitive entities with a preset sensitive word library, and a basic score is assigned according to the matching degree. By combining the frequency of occurrence and distribution density of sensitive entities, a comprehensive sensitivity feature score is obtained, specifically: S=S base +α1·f+α2·d Where S is the comprehensive sensitivity feature score, S base The base score is f, the frequency of occurrence is d, the distribution density is d, and α1 and α2 are the corresponding weights. The distribution density is obtained by dividing the number of sensitive entities by the total length of the data.

6. The data classification and grading method based on artificial intelligence according to claim 1, characterized in that, The comprehensive business feature score includes the following steps: Based on the enterprise business architecture map, identify the participation and influence of data in core business processes; By analyzing the correlation between business process nodes, the Pearson correlation coefficient between the data and key business indicators is calculated, and the absolute value of the correlation coefficient is mapped to a basic business score of 0-1. By determining the irreplaceability of the data and using redundancy to correct the basic business score, a comprehensive business feature score is obtained, specifically: B=B base *100*(1-r) Where B is the comprehensive business characteristic score, B base The score is based on basic business operations, and r represents data redundancy.

7. The data classification and grading method based on artificial intelligence according to claim 1, characterized in that, The comprehensive security risk characteristic score includes the following steps: The threat matrix analysis method is used to calculate the internal and external threat risks faced by the data. External risk dimensions include the price of data breaches in the black market and the probability of similar companies experiencing data security incidents; internal risk dimensions include the frequency of unauthorized access and the level of vulnerability in data transmission encryption strength. Each safety risk indicator is quantified into a sub-score of 0-1 points, and then a comprehensive safety risk characteristic score is obtained by weighting and summing the scores based on the probability of risk occurrence and the degree of impact. Specifically: Where R is the comprehensive security risk characteristic score, m is the total number of security risk indicators, and p j Let e ​​be the probability of occurrence of the j-th type of risk. j β represents the impact level of the j-th type of risk. j is the weight of the j-th type of risk.

8. The data classification and grading method based on artificial intelligence according to claim 1, characterized in that, The comprehensive score in step S3 includes the following steps: The overall score is as follows: Score=ω S ·S+ω B ·B+ω R ·R Where Score is the overall score, S is the overall sensitivity feature score, B is the overall business feature score, R is the overall security risk feature score, and ω is the overall security risk feature score. S ω B and ω R These are the weight coefficients corresponding to the scores of the three types of features, and the weight coefficients are dynamically generated by DAHP. Data is classified into levels and categories based on the comprehensive score, including: if the comprehensive score is ≤30, the data is level 1, corresponding to public data; if 30 < comprehensive score ≤60, the data is level 2, corresponding to internal data; if 60 < comprehensive score ≤80, the data is level 3, corresponding to sensitive data; if the comprehensive score is >80, the data is level 4, corresponding to highly sensitive data. If the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level will be directly classified as high-sensitivity data, corresponding to level 4.

9. A data classification and grading method based on artificial intelligence according to claim 1, characterized in that, Step S6 includes the following steps: The coefficient for the change in business requirements is specifically as follows: Where c1 is the business demand variation coefficient, v current v is the current business feature vector. history This is a historical business feature vector; The aforementioned safety risk change coefficient is specifically: Where c2 is the safety risk change coefficient, x current，i Let x be the value of the current i-th security risk indicator. base，i Let m be the baseline value of the i-th security risk indicator, and m be the total number of security risk indicators. The adaptive threshold is specifically: T = 0.5 + 0.3 * max(c1, c2) Where T is the adaptive threshold, c1 is the business demand change coefficient, and c2 is the security risk change coefficient; The retraining conditions for the model are: c1+c2>T and the proportion of new data ≥30%, wherein the new data must include data types added within the last 3 months.

10. A data classification and grading system based on artificial intelligence, characterized in that, Includes the following modules: Multi-source data cleaning and structuring module: Acquires multi-source data from within the enterprise and performs data cleaning and standardization processing; Cross-modal feature extraction and fusion module: Constructs a multimodal data feature extraction model, extracts multi-source features of data through a deep learning network, and generates a unified 512-dimensional global feature vector; Dynamic weight calculation and comprehensive score module: Combining dynamic hierarchical analysis and fuzzy comprehensive evaluation, it calculates the dynamic weights of sensitivity, business requirements and security risks, and calculates the comprehensive sensitivity feature score, comprehensive business feature score and comprehensive security risk feature score, outputs the comprehensive score, and classifies the data level and category according to the comprehensive score; Sensitivity Priority Rule Determination Module: Set sensitivity priority rules. When the comprehensive sensitivity feature score is greater than the preset sensitivity threshold, the data level is directly classified as high-sensitivity data, corresponding to level 4. Hierarchical security policy generation module: Generates hierarchical security policies based on data classification and grading results. The policies include differentiated configurations of data storage encryption strength, access permission granularity, and transmission encryption protocols. Dynamic monitoring and model update triggering module: Triggers model updates through dynamic monitoring mechanism: Calculates the business requirement change coefficient and the security risk change coefficient. If the sum of the two coefficients is greater than a preset threshold, the model retraining is started. The retraining adopts an incremental learning framework.

Images