Sensitive data identification and classification method, system and equipment of power monitoring system and medium

Abstract

The invention discloses a sensitive data recognition and classification method, system and device for a power monitoring system and a medium, and the method comprises the steps: obtaining text data in the power monitoring system, and carrying out the word segmentation processing of the text data, and obtaining text word segmentation data; performing semantic feature analysis through the text word segmentation data to obtain target semantic features; constructing a sensitive data recognition model through data oversampling, inputting the target semantic features into the sensitive data recognition model for sensitive data recognition, and obtaining target sensitive data; performing risk assessment on the target sensitive data based on data flow direction and data leakage risk analysis to obtain risk assessment data; and determining a security demand level of each target sensitive data according to the risk assessment data, and classifying the target sensitive data according to the security demand levels. According to the method, through two-section word segmentation processing and multi-dimensional semantic feature analysis, the accuracy and integrity of sensitive data recognition are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of sensitive data identification technology, and in particular to a method, system, device and medium for sensitive data identification and classification in a power monitoring system. Background Technology

[0002] With the increasing digitalization of power systems, power monitoring systems carry a large amount of sensitive information related to grid operation, user information, and equipment data. The security of this data is directly related to the stable operation of the power system and the protection of user privacy. To ensure the security of sensitive data during transmission, storage, and use, accurate identification and effective classification of sensitive data in power monitoring systems has become a core aspect of data security management. Currently, sensitive data identification and classification technologies mainly include rule-based dictionary matching methods and machine learning-based pattern recognition methods. These technologies can, to a certain extent, achieve automated identification and classification of sensitive data.

[0003] However, existing technologies have two main drawbacks: First, the accuracy of sensitive data identification is not high. Traditional dictionary matching methods are prone to missed detections, while machine learning-based methods often suffer from performance degradation due to imbalanced training samples and have limited accuracy in text segmentation and semantic understanding, making it difficult to accurately identify sensitive information in complex contexts. Second, the reliability of sensitive data classification is insufficient. Existing methods mainly classify data based on data type, failing to fully consider the impact of factors such as the actual flow of data and leakage risks on the security level of sensitive data. This results in classification results that cannot accurately reflect the true risk level of the data, affecting the effectiveness of subsequent security protection measures. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method, system, device, and medium for identifying and classifying sensitive data in a power monitoring system to address the problems of low accuracy in identifying sensitive data, the tendency of traditional dictionary matching methods to miss detections, and the fact that existing methods mainly rely on simple classification based on data type, resulting in classification results that cannot accurately reflect the true risk level of the data.

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

[0007] In a first aspect, the present invention provides a method for identifying and classifying sensitive data in a power monitoring system, comprising: acquiring text data from the power monitoring system; performing word segmentation on the text data to obtain word segmentation data; performing semantic feature analysis on the word segmentation data to obtain target semantic features; constructing a sensitive data identification model through data oversampling; inputting the target semantic features into the sensitive data identification model to identify sensitive data and obtain target sensitive data; performing risk assessment on each of the target sensitive data based on data flow and data leakage risk analysis to obtain risk assessment data; determining the security requirement level of each of the target sensitive data according to the risk assessment data; and classifying the target sensitive data according to the security requirement level.

[0008] As a preferred embodiment of the sensitive data identification and classification method for the power monitoring system described in this invention, the step of performing word segmentation on the text data includes: obtaining a target dictionary tree; performing preliminary word segmentation on the text data using the target dictionary tree to obtain preliminary word segmentation vocabulary; removing the preliminary word segmentation vocabulary from the text data to obtain the remaining data; performing word grouping probability analysis on the remaining data using a word segmentation model; segmenting words with a word grouping probability greater than a preset threshold to obtain the text word segmentation data.

[0009] The beneficial effects of this preferred technical solution are as follows: the two-stage word segmentation method, which performs preliminary word segmentation using the target dictionary tree and then uses a word segmentation model to perform probability analysis on the remaining data, can effectively combine the accuracy of dictionary matching with the flexibility of machine learning models.

[0010] As a preferred embodiment of the sensitive data identification and classification method for the power monitoring system described in this invention, the step of performing semantic feature analysis using text segmentation data includes: identifying the part-of-speech of each text segment in the text segmentation data; embedding the text segmentation data into text segmentation vectors based on the part-of-speech; performing lexical analysis and syntactic structure analysis on the text segmentation vectors using a syntax tree to obtain lexical data and syntactic structure data; performing contextual analysis using the text segmentation vectors through a self-learning mechanism to obtain contextual features; and obtaining the target semantic features from the lexical data, syntactic structure data, and contextual features.

[0011] As a preferred embodiment of the sensitive data identification and classification method for the power monitoring system described in this invention, the step of constructing a sensitive data identification model through data oversampling includes: acquiring training sample texts and their corresponding label data to form a sample dataset; determining the decision boundary of the sample dataset using a regression classification algorithm, and taking the minority class samples within the decision boundary range as the boundary text set; analyzing several nearest neighbors of the boundary text set, determining the sampling ratio based on the number of minority class samples, generating new minority class samples between the target nearest neighbors and the boundary text set to form a target sample dataset; and training a deep convolutional neural network model using the target sample dataset to obtain the sensitive data identification model.

[0012] As a preferred embodiment of the sensitive data identification and classification method for the power monitoring system described in this invention, the step of risk assessment of the target sensitive data includes: analyzing the transfer object according to the flow direction label of the target sensitive data, obtaining historical interaction data between the transfer object and the power monitoring system, determining the security index of the transfer object based on the frequency of feedback information and the data security transmission ratio; obtaining the data type of each target sensitive data through a type tree, and analyzing the leakage loss risk index based on the historical risk events of each data type.

[0013] As a preferred embodiment of the sensitive data identification and classification method for the power monitoring system described in this invention, the step of obtaining risk assessment data includes: evaluating the sensitivity coefficient of each target sensitive data through a graph neural network and an attention mechanism; multiplying the sensitivity coefficient, the leakage loss risk index, and the security index of the transfer object with their corresponding weight coefficients, and adding the products to obtain the risk assessment index, which is used as the risk assessment data.

[0014] The beneficial effects of this preferred technical solution are: by weighting and comprehensively calculating the risk factors of three dimensions—sensitivity coefficient, leakage loss risk index, and transfer object security index—it is possible to more comprehensively and accurately quantify the actual risk level of sensitive data.

[0015] As a preferred embodiment of the sensitive data identification and classification method for the power monitoring system described in this invention, the step of evaluating the sensitivity coefficient of the target sensitive data includes: converting the target sensitive data into a graph structure representation through a graph construction module; performing multi-layer graph convolution processing on the graph structure representation through a graph neural network layer to initially extract node features; applying an attention mechanism to the node features output by the graph neural network from both node and edge dimensions through an attention mechanism layer; integrating the attention-weighted node and edge features through a feature fusion and output module, and converting the fused features into sensitivity coefficient output through a fully connected layer.

[0016] The beneficial effects of this preferred technical solution are as follows: by adopting a hybrid architecture of graph construction module, graph neural network layer, attention mechanism layer and feature fusion and output module, it is possible to simultaneously capture the structured relationships and important feature weights of sensitive data.

[0017] Secondly, the present invention provides a sensitive data identification and classification system for a power monitoring system, comprising:

[0018] The data acquisition and word segmentation module is used to acquire text data from the power monitoring system and perform word segmentation on the text data to obtain text segmentation data.

[0019] The semantic feature analysis module is used to perform semantic feature analysis on the text segmentation data to obtain target semantic features;

[0020] The sensitive data identification module is used to construct a sensitive data identification model through data oversampling, and input the target semantic features into the sensitive data identification model to identify sensitive data and obtain the target sensitive data.

[0021] The risk assessment module is used to conduct risk assessments on the sensitive data of each target based on data flow and data leakage risk analysis, and to obtain risk assessment data.

[0022] The classification module is used to determine the security requirement level of each of the target sensitive data based on the risk assessment data, and to classify the target sensitive data according to the security requirement level.

[0023] Thirdly, the present invention provides an electronic device, comprising:

[0024] Memory and processor;

[0025] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the sensitive data identification and classification method of the power monitoring system.

[0026] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the sensitive data identification and classification method of the power monitoring system.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] By employing a two-stage word segmentation process and multi-dimensional semantic feature analysis, the accuracy and completeness of sensitive data identification are effectively improved. The method combining a target dictionary tree with a word segmentation model ensures rapid identification of common sensitive words while handling new and compound words in complex contexts, overcoming the shortcomings of traditional dictionary matching methods that are prone to missed detections. Simultaneously, data oversampling technology addresses the imbalanced training sample problem, and combining it with a deep convolutional neural network model for sensitive data identification improves the model's generalization ability and recognition accuracy, avoiding false positives and false negatives caused by sample skewness.

[0029] A multi-dimensional risk assessment system based on data flow, leakage risk, and sensitivity coefficients provides a more comprehensive reflection of the actual risk status of sensitive data compared to existing methods that simply classify data based on data type. By employing a hybrid architecture of graph neural networks and attention mechanisms to assess sensitivity coefficients, combined with the security index of the data being transferred and historical risk event analysis, precise quantification of sensitive data risk is achieved. Based on the risk assessment data, security requirement levels are determined and categorized, enabling sensitive data of different risk levels to receive corresponding security protection measures, thus improving the scientific rigor and targeted nature of data security management in power monitoring systems. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0031] Figure 1 This is a schematic diagram of the overall process of a sensitive data identification and classification method for a power monitoring system according to an embodiment of the present invention. Detailed Implementation

[0032] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0033] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for identifying and classifying sensitive data in a power monitoring system is provided, comprising S1 to S5:

[0034] S1. Obtain text data from the power monitoring system, perform word segmentation on the text data, and obtain text segmentation data.

[0035] S2. Perform semantic feature analysis on text segmentation data to obtain target semantic features.

[0036] S3. Construct a sensitive data identification model by oversampling data, input the target semantic features into the sensitive data identification model to identify sensitive data, and obtain the target sensitive data.

[0037] S4. Based on data flow and data leakage risk analysis, conduct risk assessments on the sensitive data of each target to obtain risk assessment data.

[0038] S5. Determine the security requirement level of each target's sensitive data based on the risk assessment data, and classify the target's sensitive data according to the security requirement level.

[0039] It should be noted that power monitoring systems carry a large amount of sensitive information related to power grid operation, user privacy, and equipment parameters. The security of this data is directly related to the stable operation of the power system and data security. In practical applications, traditional sensitive data identification methods mainly rely on manual identification and dictionary matching. Manual identification depends heavily on the professional competence of relevant personnel, and different personnel may have different standards for identifying sensitive data, leading to discrepancies in the identification results. While dictionary matching methods can ensure consistency, the identification accuracy is low, and if the data dictionary is incomplete, sensitive data may be missed. At the same time, existing sensitive data classification methods usually only classify data based on data type, failing to fully consider the impact of factors such as the actual flow of data and the risk of leakage on the security level of sensitive data. This results in classification results that cannot accurately reflect the true risk level of the data, affecting the pertinence and effectiveness of subsequent security protection measures.

[0040] Therefore, to address the aforementioned issues of identification accuracy and classification reliability, steps S1-S5 are employed, using a two-stage word segmentation process combined with deep semantic analysis to improve the word segmentation accuracy and semantic understanding capabilities of text data. Oversampling technology is used to resolve the imbalanced training sample problem, constructing a high-precision sensitive data identification model to achieve accurate identification of sensitive data. A multi-dimensional risk assessment system based on data flow, leakage loss risk, and sensitivity coefficients is established. Graph neural networks and attention mechanisms are used to assess data sensitivity, and weighted comprehensive calculations yield scientific risk assessment results. Ultimately, this achieves accurate classification of sensitive data based on risk levels, providing reliable assurance for data security management in power monitoring systems.

[0041] Example 2, refer to Figure 1 As an embodiment of the present invention, based on the above embodiment, a method for identifying and classifying sensitive data in a power monitoring system is provided.

[0042] In this embodiment of the application, step S1, the step of segmenting the text data, includes A1 to A3:

[0043] A1. Obtain the target dictionary tree, and perform preliminary word segmentation on the text data using the target dictionary tree to obtain preliminary word segmentation vocabulary;

[0044] Specifically, the process begins by collecting text data from the power monitoring system, including user operation text data and text data that needs to be uploaded to the system. The target dictionary tree uses a tree structure to store predefined sensitive and commonly used words. These words are pre-entered by relevant professionals based on the business characteristics of the power monitoring system. When performing text matching through the target dictionary tree, the longest match principle is used. Starting from the beginning of the text, each character is scanned sequentially to find the longest matching word in the dictionary tree as the segmentation result, thus completing the initial segmentation process for the entire text.

[0045] A2. Remove the initial segmented words from the text data to obtain the remaining data;

[0046] The original text data is labeled. The initial segmented words successfully matched in step A1 are marked with their positions in the original text, recording the start and end positions of each segmented word. Then, based on the position information, the labeled initial segmented words are removed from the original text data, extracting continuous text segments that have not undergone segmentation. During the removal process, the relative positional relationships and contextual continuity of the remaining text segments are maintained. All remaining unsegmented text segments are reorganized to form the remaining dataset for further processing.

[0047] A3. Perform word grouping probability analysis on the remaining data using a word segmentation model, segment words with a word grouping probability greater than a preset threshold, and obtain text word segmentation data.

[0048] The word segmentation model is built using machine learning algorithms. A training dataset is created by collecting sample text and its corresponding labels, where the labels indicate whether a text fragment constitutes a valid word. The machine learning model is trained using this training dataset to learn the word formation patterns and probabilistic features of different text fragments. After training, this model is applied to the remaining text data. The word segmentation model analyzes the contextual information of consecutive characters in the remaining text data to calculate the probability of them forming words. When the probability exceeds a preset threshold, the text fragment is identified as a valid word and segmented.

[0049] It should be noted that step A2 ensures the orderly execution of the two-stage word segmentation process. By accurately removing processed words, it avoids the problems of repeated word segmentation and word segmentation conflicts. At the same time, it provides the word segmentation model with clean data to be processed, improving the accuracy and efficiency of subsequent machine learning word segmentation.

[0050] In one optional implementation, the word segmentation of text data in step S1 may also include post-processing optimization of the word segmentation results. The reasonableness of the word segmentation results is verified by part-of-speech tagging and grammatical rules, and obviously unreasonable word segmentation results are corrected to further improve the overall quality of word segmentation and provide more reliable basic data for subsequent semantic feature analysis.

[0051] In this embodiment of the application, step S2, the step of performing semantic feature analysis using text segmentation data, includes B1 to B4:

[0052] B1. Identify the part-of-speech tag of each word in the text segmentation data, and perform word embedding on the text segmentation data according to the part-of-speech tag to obtain the text segmentation vector;

[0053] Specifically, a pre-constructed part-of-speech (POS) mapping table is used to identify the part-of-speech tags of each text segment in the text segmentation data. This mapping table includes common POS categories such as nouns, verbs, adjectives, and adverbs, as well as POS tags for professional terms used in power monitoring systems. After POS identification, corresponding word embedding strategies are applied based on different POS categories to convert each text segment into a corresponding numerical vector representation. During word embedding, noun segmentation focuses more on entity semantic features, verb segmentation focuses more on action relationship features, and adjective segmentation focuses more on attribute description features. This POS-guided embedding method yields text segmentation vectors with rich semantic information.

[0054] B2. Perform lexical and syntactic structure analysis on the text segmentation vectors using a syntax tree to obtain lexical and syntactic structure data.

[0055] The syntax tree (SMR) uses a tree-like abstract structure to represent syntactic relations, where each node corresponds to a specific syntactic structural component, such as subject, predicate, or object. The SMR is constructed using a recursive descent parsing algorithm, first identifying the core structure of the sentence, and then progressively expanding on modifiers and dependencies. During lexical analysis, the grammatical function and syntactic role information of each word segment is extracted to form lexical data. During syntactic structure analysis, dependency, modification, and coordination relationships between word segments are analyzed. By traversing each node of the SMR, complete syntactic structure data is obtained, providing a structured grammatical foundation for subsequent semantic feature extraction.

[0056] B3. Contextual features are obtained by using text segmentation vectors through a self-learning mechanism to analyze the context.

[0057] The self-learning mechanism employs an attention-based context analysis method, using a sliding window technique to acquire contextual information for each text segment, with the window size adjusted according to text complexity. It calculates relevance weights between segments at different positions using the segmented text vectors, focusing on contextual fragments with high semantic relevance to the current segment. Through a multi-layered neural network structure, it learns deep semantic connections within the context, automatically capturing implicit semantic relationships and logical dependencies between words, ultimately extracting contextual feature vectors that reflect the overall contextual characteristics of the text.

[0058] B4. Obtain the target semantic features from lexical data, syntactic structure data, and contextual features;

[0059] The generation of target semantic features employs a feature fusion strategy, comprehensively processing lexical data, syntactic structure data, and contextual features with different weights. Lexical data provides basic lexical semantic information, syntactic structure data provides grammatical relational information, and contextual features provide overall semantic environment information. Through feature concatenation and dimensionality reduction, the three types of features are integrated into a unified target semantic feature representation, ensuring that the subsequent sensitive data identification model can obtain comprehensive and accurate semantic input.

[0060] In an optional implementation, when performing semantic feature analysis in step S2, a domain knowledge base can be introduced to perform special processing on the professional terms of the power monitoring system. Through semantic enhancement and association analysis of professional terms, the quality of semantic feature extraction in the sensitive data identification scenario in the power field can be improved, further enhancing the professionalism and accuracy of the overall identification effect.

[0061] In this embodiment of the application, step S3, the step of constructing a sensitive data identification model through data oversampling, includes C1 to C4:

[0062] C1. Obtain the training sample text and its corresponding label data to form a sample dataset;

[0063] Specifically, historical text data from the power monitoring system is collected as training samples, including confirmed sensitive and non-sensitive data samples. Each training sample text is manually labeled using a binary classification method, where a label "1" indicates that the text fragment contains sensitive data, and a label "0" indicates that the text fragment does not contain sensitive data. The training sample texts are then paired with their corresponding labels to form a complete sample dataset, ensuring that the dataset contains a sufficient number of positive and negative samples for subsequent model training.

[0064] C2. Determine the decision boundary of the sample dataset through regression classification algorithm, and take the minority class samples within the decision boundary range as the boundary text set;

[0065] The Logistic Regression classification algorithm was used to analyze the sample dataset, and the probability value of each sample belonging to the sensitive data class was calculated using the sigmoid function. Based on the probability distribution, the decision boundary of the sample dataset was determined, and samples with probability values ​​between 0.3 and 0.7 were identified as boundary samples, which are typically located in the fuzzy region of classification decision. From the boundary samples, samples belonging to the minority class (sensitive data class) were selected and treated as the boundary text set. These boundary samples have high classification uncertainty and are the focus of data oversampling.

[0066] C3. Analyze several nearest neighbors of the boundary text set, determine the sampling ratio based on the number of minority class samples, generate new minority class samples between the target nearest neighbors and the boundary text set, and form the target sample dataset;

[0067] The K-nearest neighbors algorithm is used to analyze several nearest neighbor samples for each sample in the boundary text set, typically selecting 5-10 nearest neighbors for analysis. The sampling ratio is determined based on the ratio of minority class samples to majority class samples in the sample dataset. When the number of minority class samples is less than 30% of the number of majority class samples, the sampling ratio is set to 2-3 times. A predetermined number of target nearest neighbors are randomly selected from the nearest neighbors. The feature attribute difference between the target nearest neighbors and the boundary samples is calculated. This difference is multiplied by a random number between 0 and 1. Based on the calculation result and the sampling ratio, new minority class samples are generated in the feature space between the two. This process is repeated until the number of newly generated minority class samples reaches a predetermined standard. The new samples are then added to the original sample dataset to form the target sample dataset.

[0068] C4. Train the deep convolutional neural network model using the target sample dataset to obtain a sensitive data identification model;

[0069] The target sample dataset was divided into training and test sets in an 8:2 ratio, and a deep convolutional neural network model was constructed. The model's network structure parameters were set, including the number of input layer nodes corresponding to the semantic feature vector dimension, the use of a combination of multiple convolutional and pooling layers in the hidden layers, and the output layer being set to have 2 nodes corresponding to the binary classification result. The model was iteratively trained using the training set with a learning rate of 0.001 and 100 training epochs. The model's accuracy was evaluated using a validation set after each epoch. Training was stopped when the validation set accuracy failed to improve for five consecutive epochs. The final performance evaluation of the trained model was performed using the test set to verify its generalization ability and classification accuracy. The model that passed the test was selected as the sensitive data identification model.

[0070] It should be noted that the data oversampling technique effectively solves the problem of sample imbalance caused by the scarcity of sensitive data samples in the power monitoring system. By reasonably generating new minority class samples in the feature space, it improves the model's ability to identify sensitive data and avoids the problems of high false positive rate and low recall rate caused by sample skew in traditional methods.

[0071] In an optional implementation, when constructing the sensitive data identification model in step S3, an ensemble learning method can be used for model optimization. This involves training multiple deep convolutional neural network sub-models with different parameter settings and integrating the prediction results of multiple sub-models using a voting mechanism or a weighted average method. Specifically, 3-5 sub-models can be constructed, each employing a different network depth, convolutional kernel size, or activation function. Cross-validation is used to determine the weight coefficients of each sub-model, and the final output sensitive data identification result is a weighted sum of the prediction results from all sub-models. This ensemble learning method can further improve the model's robustness and prediction accuracy, especially when dealing with complex and variable text data from power monitoring systems, effectively reducing the risk of overfitting that may occur with a single model.

[0072] In this embodiment of the application, step S4, the step of risk assessment of the target sensitive data, includes D1 to D2:

[0073] D1. Analyze the flow objects based on the flow direction labels of the target sensitive data, obtain the historical interaction data between the flow objects and the power monitoring system, and determine the security index of the flow objects based on the frequency of feedback information and the data security transmission ratio.

[0074] Specifically, sensitive data in the power monitoring system all have preset flow direction labels. By parsing these labels, the specific flow path and target object of the sensitive data are determined. These targets include other business systems, external terminal devices, and third-party application interfaces. The system log database is queried to obtain historical interaction records between the flow objects and the power monitoring system over the past 12 months, extracting system-generated access logs and security event records. The frequency of access denials to flow objects is statistically analyzed, and the data security transmission success rate (the ratio of secure transmissions to total transmissions) is calculated. A security index calculation model is established based on the denial frequency and the secure transmission ratio. A higher denial frequency and a lower secure transmission ratio result in a lower security index for the flow object, and vice versa.

[0075] D2. Obtain the data type of each target's sensitive data through the type tree, and analyze the leakage loss risk index based on the historical risk events of each data type;

[0076] A type tree structure with multiple levels of nodes is constructed, with each node corresponding to a specific data type judgment rule, such as user identity information, equipment parameter data, and power grid operation data. By traversing each node of the type tree, the target sensitive data is matched with the judgment rules of the nodes to determine the specific data type of each sensitive data. A historical risk event database is queried to obtain records of security events such as data breaches, unauthorized access, and data loss that have occurred in the past for each data type. A judgment matrix is ​​established, and the risk parameter weights are determined based on factors such as the frequency of occurrence, scope of impact, and degree of loss of historical risk events. The comprehensive leakage loss risk index for each data type is then calculated.

[0077] In step S4, the steps for obtaining risk assessment data include D3 to D4:

[0078] D3. Sensitivity coefficients of sensitive data for each target are evaluated using graph neural networks and attention mechanisms;

[0079] The steps for evaluating the sensitivity coefficient of target sensitive data include D3.1 to D3.4:

[0080] D3.1. Convert the target sensitive data into a graph structure representation using the graph construction module;

[0081] In this embodiment, the graph construction module converts the target sensitive data and its related attribute information into a graph structure representation, where sensitive data entities serve as nodes in the graph, and the relationships between data are represented as edges. Node features include multi-dimensional attribute information such as semantic features, type features, and importance of the data, while edge features include relational attributes such as the strength of association between data, dependency relationships, and co-occurrence frequency. The graph structure data is stored in the form of an adjacency matrix and a feature matrix, providing a standardized input format for subsequent graph neural network processing.

[0082] D3.2. Perform multi-layer graph convolution processing on the graph structure representation through graph neural network layers to initially extract node features;

[0083] The graph neural network layer employs a multi-layer graph convolutional neural network architecture, typically using 3-5 convolutional layers for feature extraction. Each graph convolutional layer aggregates neighbor information and its own features to progressively extract higher-level node representation features. The first convolutional layer primarily extracts local neighborhood features, the intermediate convolutional layers expand the receptive field, and the final convolutional layer generates global semantic features. Residual connections and batch normalization techniques are used during the convolution process to prevent gradient vanishing and overfitting problems, ensuring the network can effectively learn complex relationship patterns in graph structures.

[0084] D3.3. Apply attention mechanisms to the node features output by the graph neural network through the attention mechanism layer, from the two dimensions of nodes and edges.

[0085] The attention mechanism layer calculates attention weights from both node and edge dimensions. The node attention mechanism calculates the importance weights between nodes through self-attention, identifying the nodes most critical for sensitivity assessment; the edge attention mechanism calculates the importance weights of edges, identifying the most important relationships for node feature propagation. Both attention weights are normalized using a softmax function to ensure a reasonable weight distribution and a sum of 1, and then weighted separately for node features and edge features.

[0086] D3.4 Integrate the attention-weighted node and edge features through the feature fusion and output module, and convert the fused features into sensitivity coefficients through a fully connected layer;

[0087] The feature fusion and output module first concatenates or adds the attention-weighted node and edge features to form a comprehensive graph feature representation. A multi-layer fully connected neural network is then used to perform non-linear transformations and dimensionality reduction on the fused features, ultimately outputting a single sensitivity coefficient value. The fully connected layers employ the ReLU activation function and Dropout regularization, while the output layer uses the sigmoid function to limit the sensitivity coefficient to the range of 0-1; the closer the value is to 1, the higher the data sensitivity.

[0088] D4. Multiply the sensitivity coefficient, leakage loss risk index, and security index of the transferred object by their respective weight coefficients, and add the products together to obtain the risk assessment index, which is used as the risk assessment data.

[0089] The sensitivity coefficient, leakage loss risk index, and transfer object security index are assigned weights of 0.4, 0.3, and 0.3, respectively, with the weight allocation determined based on the actual security requirements of the power monitoring system. Each of the three indicators is multiplied by its corresponding weight to obtain a weighted risk component. All weighted components are then summed to obtain the final risk assessment index. This index reflects the overall risk level of the target sensitive data and is used as risk assessment data for subsequent security requirement level determination and classification.

[0090] It should be noted that the multi-dimensional risk assessment method, by comprehensively considering the security of data flow, the risk characteristics of data types, and the sensitivity of the data itself, can more accurately quantify the actual risk level of sensitive data, avoid the risk assessment bias that may be caused by a single assessment dimension, and provide a scientific basis for the subsequent formulation of targeted data security protection strategies.

[0091] In an optional implementation, during risk assessment in step S4, a risk adjustment mechanism can be introduced to adjust the weight coefficients of each risk assessment indicator by real-time monitoring of changes in the security status of the power monitoring system. Specifically, when the system detects an increase in a specific type of security threat, the weight of the leakage loss risk index is increased accordingly; when the security performance of certain transferred objects is found to have declined, the weight of the transferred object's security index is increased; and when regulations and policies related to data sensitivity change, the weight of the sensitivity coefficient is adjusted. By establishing a risk weight adjustment table, the weight configuration is automatically updated according to real-time changes in the system's security status, enabling the risk assessment results to more accurately reflect the current actual risk situation and improving the adaptability and practicality of the entire sensitive data classification system.

[0092] In this embodiment, step S5: Determine the security requirement level of each target sensitive data based on the risk assessment data, and classify the target sensitive data according to the security requirement level.

[0093] Specifically, a database mapping risk assessment data to security requirement levels is established, categorizing risk assessment indices into different security requirement levels based on their numerical range. A risk assessment index between 0 and 0.3 corresponds to a low security requirement level, indicating a relatively low risk of leakage for the sensitive data, requiring basic security measures. A risk assessment index between 0.3 and 0.6 corresponds to a medium security requirement level, indicating a medium risk of leakage for the sensitive data, requiring standard security measures. A risk assessment index between 0.6 and 0.8 corresponds to a high security requirement level, indicating a relatively high risk of leakage for the sensitive data, requiring enhanced security measures. A risk assessment index between 0.8 and 1.0 corresponds to an extremely high security requirement level, indicating an extremely high risk of leakage for the sensitive data, requiring the highest level of security measures.

[0094] Through a database matching mechanism, the risk assessment index of each target sensitive data is compared with a preset level threshold to automatically determine its corresponding security requirement level. The matching process employs a range-based algorithm, first determining which numerical range the risk assessment index belongs to, and then directly mapping it to the corresponding security requirement level. For sensitive data at boundary values, a rounding-up principle is used; that is, when the risk assessment index is exactly equal to the range boundary value, it is classified into a higher security requirement level to ensure adequate security protection.

[0095] Sensitive data is categorized and classified according to the determined security requirement levels, grouping sensitive data with the same security requirement level into the same category. Four classification containers are established, corresponding to low, medium, high, and extremely high security requirement levels, respectively. Sensitive data is stored in the corresponding classification container according to its security requirement level. Each classification container contains detailed information on all sensitive data under that level, including complete attribute information such as data content, data type, transfer object, and risk assessment index.

[0096] Develop corresponding data processing strategies and security protection measures for each security requirement level. Sensitive data with low security requirements is transmitted with conventional encryption and basic access control; sensitive data with medium security requirements is transmitted with enhanced encryption algorithms and authentication mechanisms; sensitive data with high security requirements is transmitted with multiple encryptions, digital signatures, and strict access auditing; sensitive data with extremely high security requirements is transmitted with comprehensive security measures such as end-to-end encryption, multi-factor authentication, real-time monitoring, and offline storage.

[0097] It should be noted that the hierarchical classification method based on risk assessment data can formulate differentiated security protection strategies according to the actual risk level of sensitive data, avoiding the waste of resources or insufficient protection that may be caused by applying uniform protection measures to all sensitive data in traditional methods, and realizing the refinement and efficiency of sensitive data security management.

[0098] In an optional implementation, during the sensitive data classification in step S5, a classification adjustment mechanism can be established. This mechanism involves periodically reassessing the risk status of sensitive data and adjusting its security requirement level and classification. Specifically, a monthly or quarterly assessment cycle is set, and the risk assessment index of sensitive data is recalculated. When a change in the risk assessment index of a sensitive data point is detected, its security requirement level is automatically adjusted and it is moved to the corresponding classification container. Simultaneously, a classification change log is established to record the time, reason, and content of each level adjustment, providing security managers with historical tracking information on classification changes. This adjustment mechanism ensures that the sensitive data classification results always reflect the latest risk status, improving the timeliness and accuracy of the entire data security management system.

[0099] In summary, the two-stage word segmentation process and multi-dimensional semantic feature analysis effectively improve the accuracy and completeness of sensitive data identification. The method combining a target dictionary tree with a word segmentation model ensures rapid identification of common sensitive words while handling new and compound words in complex contexts, overcoming the shortcomings of traditional dictionary matching methods that are prone to missed detections. Furthermore, the use of data oversampling technology addresses the imbalanced training sample problem, and the integration of a deep convolutional neural network model for sensitive data identification improves the model's generalization ability and recognition accuracy, avoiding false positives and false negatives caused by sample skewness.

[0100] A multi-dimensional risk assessment system based on data flow, leakage risk, and sensitivity coefficients provides a more comprehensive reflection of the actual risk status of sensitive data compared to existing methods that simply classify data based on data type. By employing a hybrid architecture of graph neural networks and attention mechanisms to assess sensitivity coefficients, combined with the security index of the data being transferred and historical risk event analysis, precise quantification of sensitive data risk is achieved. Based on the risk assessment data, security requirement levels are determined and categorized, enabling sensitive data of different risk levels to receive corresponding security protection measures, thus improving the scientific rigor and targeted nature of data security management in power monitoring systems.

[0101] Example 3 illustrates a schematic scheme for a sensitive data identification and classification method in a power monitoring system. It should be noted that the technical solution of this sensitive data identification and classification system in the power monitoring system is based on the same concept as the aforementioned sensitive data identification and classification method in the power monitoring system. Details not described in detail in this embodiment can be found in the description of the aforementioned sensitive data identification and classification method in the power monitoring system.

[0102] This embodiment also provides a sensitive data identification and classification system for a power monitoring system, including:

[0103] The data acquisition and word segmentation module is used to acquire text data from the power monitoring system and perform word segmentation on the text data to obtain text segmentation data.

[0104] The semantic feature analysis module is used to perform semantic feature analysis on the text segmentation data to obtain target semantic features;

[0105] The sensitive data identification module is used to construct a sensitive data identification model through data oversampling, and input the target semantic features into the sensitive data identification model to identify sensitive data and obtain the target sensitive data.

[0106] The risk assessment module is used to conduct risk assessments on the sensitive data of each target based on data flow and data leakage risk analysis, and to obtain risk assessment data.

[0107] The classification module is used to determine the security requirement level of each of the target sensitive data based on the risk assessment data, and to classify the target sensitive data according to the security requirement level.

[0108] This embodiment also provides an electronic device suitable for sensitive data identification and classification in power monitoring systems, comprising: 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 sensitive data identification and classification method for power monitoring systems as proposed in the above embodiments.

[0109] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the sensitive data identification and classification method for a power monitoring system as proposed in the above embodiments.

[0110] The storage medium proposed in this embodiment and the sensitive data identification and classification method for power monitoring systems proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0111] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0112] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not 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 identifying and classifying sensitive data in a power monitoring system, characterized in that, include: Text data from a power monitoring system is acquired, and the text data is segmented into words to obtain segmented text data. Semantic feature analysis is performed on the text segmentation data to obtain the target semantic features; A sensitive data identification model is constructed by oversampling data. The target semantic features are then input into the sensitive data identification model to identify sensitive data and obtain the target sensitive data. Based on data flow and data leakage risk analysis, risk assessment is performed on the sensitive data of each target to obtain risk assessment data; Based on the risk assessment data, the security requirement level of each of the target sensitive data is determined, and the target sensitive data is classified according to the security requirement level.

2. The sensitive data identification and classification method for a power monitoring system as described in claim 1, characterized in that, The steps for word segmentation of text data include: Obtain the target dictionary tree, and perform preliminary word segmentation on the text data using the target dictionary tree to obtain preliminary word segmentation vocabulary; The remaining data is obtained by removing initial word segmentation words from the text data; The remaining data is analyzed for word grouping probability using a word segmentation model. Words with a word grouping probability greater than a preset threshold are segmented to obtain the text segmentation data.

3. The sensitive data identification and classification method for a power monitoring system as described in claim 2, characterized in that, The steps for semantic feature analysis using text segmentation data include: Identify the part-of-speech tag of each text segment in the text segmentation data, and perform word embedding on the text segmentation data according to the part-of-speech tag to obtain the text segmentation vector; Lexical and syntactic structure analysis are performed on the word segmentation vectors of the text using a syntax tree to obtain lexical data and syntactic structure data. By using the text segmentation vectors through a self-learning mechanism, contextual features are obtained through contextual analysis. The target semantic features are obtained from the lexical data, syntactic structure data, and contextual features.

4. The sensitive data identification and classification method for a power monitoring system as described in claim 3, characterized in that, The steps involved in building a sensitive data identification model through data oversampling include: Obtain the training sample text and its corresponding label data to form a sample dataset; The decision boundary of the sample dataset is determined by a regression classification algorithm, and the minority class samples within the decision boundary range are taken as the boundary text set. Analyze several nearest neighbors of the boundary text set, determine the sampling ratio based on the number of minority class samples, and generate new minority class samples between the target nearest neighbors and the boundary text set to form the target sample dataset; The sensitive data identification model is obtained by training a deep convolutional neural network model using the target sample dataset.

5. The sensitive data identification and classification method for a power monitoring system as described in claim 4, characterized in that, The steps for conducting a risk assessment on the target sensitive data include: The flow direction of the target sensitive data is analyzed to identify the flow object, and historical interaction data between the flow object and the power monitoring system is obtained. The security index of the flow object is determined based on the frequency of feedback information and the data security transmission ratio. The data types of the target sensitive data are obtained through a type tree, and the leakage loss risk index is analyzed based on the historical risk events of each data type.

6. The sensitive data identification and classification method for a power monitoring system as described in claim 5, characterized in that, The steps to obtain risk assessment data include: Sensitivity coefficients for the target sensitive data are evaluated using graph neural networks and attention mechanisms. The sensitivity coefficient, leakage loss risk index, and security index of the transferred object are multiplied by their respective weight coefficients, and the products are added together to obtain the risk assessment index, which is used as the risk assessment data.

7. The sensitive data identification and classification method for a power monitoring system as described in claim 6, characterized in that, The steps for evaluating the sensitivity coefficient of the target sensitive data include: The target sensitive data is converted into a graph structure representation using the graph construction module; The graph structure representation is processed by multi-layer graph convolution through graph neural network layers to initially extract node features; The attention mechanism layer applies attention to the node features output by the graph neural network from two dimensions: nodes and edges. The feature fusion and output module integrates attention-weighted node and edge features, and the fused features are converted into sensitivity coefficients for output through a fully connected layer.

8. A sensitive data identification and classification system for a power monitoring system, using the method described in any one of claims 1-7, characterized in that, include: The data acquisition and word segmentation module is used to acquire text data from the power monitoring system and perform word segmentation on the text data to obtain text segmentation data. The semantic feature analysis module is used to perform semantic feature analysis on the text segmentation data to obtain target semantic features; The sensitive data identification module is used to construct a sensitive data identification model through data oversampling, and input the target semantic features into the sensitive data identification model to identify sensitive data and obtain the target sensitive data. The risk assessment module is used to conduct risk assessments on the sensitive data of each target based on data flow and data leakage risk analysis, and to obtain risk assessment data. The classification module is used to determine the security requirement level of each of the target sensitive data based on the risk assessment data, and to classify the target sensitive data according to the security requirement level.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the sensitive data identification and classification method of the power monitoring system according to any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the sensitive data identification and classification method of the power monitoring system according to any one of claims 1 to 7.

Images