A method and system for semantic-level sensitive data identification and dynamic desensitization for large-scale power models

By employing cross-modal feature fusion and dynamic desensitization methods, the problems of low coverage and high misjudgment in power data identification are solved, achieving high-precision sensitive data identification and desensitization, ensuring data logical consistency, and applicable to multi-source heterogeneous data processing in power systems.

CN123069412APending Publication Date: 2026-11-10GUIZHOU POWER GRID CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611026197.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-10
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing methods for desensitizing sensitive power information suffer from low recognition coverage, high false positive rates due to ambiguous labels, and the inability of desensitized data to be used for subsequent power analysis due to its separation from the original data logic. Furthermore, they fail to deeply mine hidden semantics, resulting in blind spots in recognition and frequent false detections.

Method used

By acquiring multi-source heterogeneous data from the power system, cross-modal features are extracted in parallel and mapped to a high-order tensor product space for implicit semantic association fusion. Sensitive data is identified by combining conditional mutual information prior matrix constraints, and dynamic desensitization is performed through content topology consistency loss and privacy dual discrimination constraints. Finally, the desensitization risk is quantitatively assessed and adaptive closed-loop update is performed.

Benefits of technology

It achieves high-precision semantic-level sensitive data identification, ensuring that de-identified data retains the original business characteristics and logic, reducing the probability of false detection and missed detection, applicable to the protection of different sensitive data types, reducing operation and maintenance costs, and adapting to the power system architecture without the need to superimpose other systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069412A_ABST
    Figure CN123069412A_ABST
Patent Text Reader

Abstract

This invention discloses a semantic-level sensitive data identification and dynamic desensitization method and system for large-scale power models, relating to the field of power data processing. The method acquires multi-source heterogeneous power data from a power system and performs parallel cross-modal feature extraction. Multimodal sensitive features are mapped to a high-order tensor product space for implicit semantic association fusion to generate a unified high-dimensional feature set. Under the constraint of a conditional mutual information prior matrix, these features are mapped one by one into a continuous label distribution vector, and the membership degree of the vector sensitive labels is calculated for sensitive data identification. Sensitive data is subjected to content topology consistency loss constraints and privacy dual-discrimination constraints. Dynamic desensitization is performed on the relevant sensitive data to obtain desensitized power data. Finally, the desensitization mask is dynamically adjusted and the desensitized power data is updated adaptively in a closed loop through a desensitization risk quantification assessment to obtain the final desensitized power data. This application achieves closed-loop protection with high-precision semantic recognition and high-fidelity dynamic desensitization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the technical field of power data processing, and in particular to a method and system for semantic-level sensitive data identification and dynamic desensitization for large power models. Background Technology

[0002] Currently, with the in-depth construction of new power systems, the application of artificial intelligence big data models in core business scenarios such as power dispatching, power plant operation and maintenance, customer service, and power grid planning is becoming increasingly widespread. The data processed by the models covers massive amounts of sensitive information such as users' personal privacy information, core power operation data, and key equipment information. This data is related to the safe and stable operation of the power grid, the legitimate rights and interests of users, and national energy security. Once the information is leaked or tampered with, it may lead to serious consequences such as power grid dispatching errors, equipment failures, user privacy leaks, and energy supply interruptions. Therefore, it is necessary to de-identify the sensitive information in the power data.

[0003] Existing methods for desensitizing sensitive power information typically involve identifying sensitive data based on rule matching and using fixed masks for static desensitization. However, power data is highly heterogeneous, including both structured and unstructured data. Separating and extracting relevant information according to data type and then associating entities suffers from cross-modal fragmentation, low recognition coverage, and is prone to blind spots. Furthermore, the strong coupling and ambiguity between power grid operation data lead to significant fuzziness and multiple causal relationships in the corresponding data labels. Data identification by labeling can only determine the existence of labels, but cannot quantify the coupling or association between multiple labels. Especially in complex and ambiguous contexts such as synonym replacement and multi-business intersections, the inability to deeply mine hidden semantics results in frequent missed and false detections. Existing rule-based character masking or synonym replacement can destroy the feature distribution, statistical regularity, and business context logic of the original power data, resulting in significant differences between the desensitized data and the original business data, making it unusable for subsequent power analysis. Summary of the Invention

[0004] To address the problems of low sensitive data identification coverage, high misjudgment due to label ambiguity, and desensitized data being detached from the original data logic and unusable for subsequent power analysis in existing technologies, this application provides a semantic-level sensitive data identification and dynamic desensitization method and system for large power models, which can achieve closed-loop protection of high-precision semantic identification and high-fidelity dynamic desensitization.

[0005] Firstly, the aforementioned inventive objective of this application is achieved through the following technical solution: A semantic-level sensitive data identification and dynamic desensitization method for large-scale power models, the method comprising: Multi-source heterogeneous power data from the power system is acquired and cross-modal features are extracted in parallel. The extracted multi-modal sensitive features are mapped to a high-order tensor product space for implicit semantic association fusion processing to obtain a unified high-dimensional feature set. Under the constraint of the conditional mutual information prior matrix, the unified high-dimensional feature set is mapped one by one to a continuous label distribution vector and the membership degree of the vector sensitive label is calculated to identify sensitive data and obtain the sensitive data identification result. The sensitive data identification results are subjected to content topology consistency loss constraint and privacy dual discrimination constraint processing respectively. Based on the constraint results, the relevant sensitive data are dynamically desensitized to obtain power desensitized data. The power desensitized data is subjected to a quantitative assessment of desensitization risk. Based on the assessment results, the desensitization mask is dynamically adjusted and the power desensitized data is updated adaptively in a closed loop to obtain the final power desensitized data.

[0006] In a preferred embodiment, this application can be further configured as follows: The acquisition of multi-source heterogeneous power data from the power system and the parallel extraction of cross-modal features, followed by mapping the extracted multi-modal sensitive features to a high-order tensor product space for implicit semantic association fusion processing to obtain a unified high-dimensional feature set, specifically includes: The structured data in the multi-source heterogeneous power data is standardized and preprocessed to obtain structured features, and the unstructured data in the multi-source heterogeneous power data is extracted in parallel using multimodal features to obtain unstructured features. The structured and unstructured features are mapped to a pre-constructed high-order tensor product space for feature cross-attention fusion processing to construct a feature fusion matrix; Based on the feature fusion matrix, tensor multiplication is performed on the structured features and the unstructured features to obtain a unified high-dimensional feature set.

[0007] In a preferred embodiment, this application can be further configured as follows: mapping the structured and unstructured features to a pre-constructed high-order tensor product space for feature cross-attention fusion processing to construct a feature fusion matrix specifically includes: The expression for the feature fusion matrix is ​​as follows: (1) in, Represents the feature fusion matrix. , These represent the query projection matrix and the key projection matrix of the cross-attention mechanism, respectively. Indicates the cross-attention head dimension. , These represent structured features and unstructured features, respectively.

[0008] In a preferred embodiment, this application can be further configured as follows: the tensor product operation on the structured features and the unstructured features based on the feature fusion matrix to obtain a unified high-dimensional feature set specifically includes: The tensor product expression for the structured features and the unstructured features is as follows: (2) in, Represents unified high-dimensional features. The value projection matrix represents the cross-attention mechanism. Represents the feature fusion matrix. This represents the tensor product operation.

[0009] In a preferred embodiment, this application can be further configured as follows: under the constraint of the conditional mutual information prior matrix, mapping the unified high-dimensional feature set one by one to a continuous label distribution vector and calculating the membership degree of the vector sensitive labels to perform sensitive data identification, thereby obtaining the sensitive data identification result, specifically includes: Under the constraint of label probability distribution, the unified high-dimensional feature set is mapped one by one to a continuous label distribution vector, and the expression of the continuous label distribution vector is as follows: (3) in, Indicates the first A continuous label distribution vector of uniform high-dimensional features from each input. to They represent the first Tag For the first A unified high-dimensional feature The probability of membership. This represents the first in the power sensitive data tagging system. A specific label; The expression for the tag probability distribution constraint is as follows: (4) The continuous label distribution vector is constrained by a conditional mutual information prior matrix, and the asymmetric divergence loss value of the predicted label distribution is calculated during the feature label mapping process to obtain the comprehensive loss coefficient. Based on the comprehensive loss coefficient, the mapping parameters of the continuous label distribution vector are back-optimized and the membership degree of sensitive labels is calculated. When the membership degree of the sensitive label exceeds the preset sensitivity judgment threshold, it is marked as sensitive data, and the sensitive data identification result is obtained.

[0010] In a preferred embodiment, this application can be further configured as follows: the process of constraining the continuous label distribution vector with a conditional mutual information prior matrix and calculating the asymmetric divergence loss value of the predicted label distribution during feature label mapping to obtain the conditional mutual information prior matrix constraint in the comprehensive loss coefficient specifically includes: The expression for the conditional mutual information prior matrix constraint is as follows: (5) in, Indicates conditional mutual information. Represents environment variables for business scenarios. This refers to specific environment variables within the business scenario environment variables. Representing environment variables The prior probability of occurrence, These represent environment variables. Next tag ,Label The conditional probability of occurrence Representing environment variables Next tag ,Label Joint probability of co-occurrence.

[0011] In a preferred embodiment, this application can be further configured as follows: the process of constraining the continuous label distribution vector with a conditional mutual information prior matrix and calculating the asymmetric divergence loss value of the predicted label distribution during feature label mapping to obtain the asymmetric divergence loss value in the comprehensive loss coefficient specifically includes: The expression for calculating the asymmetric divergence loss value is as follows: (6) in, Represents the asymmetric divergence value. This represents the number of features in a unified high-dimensional feature set. This indicates the number of tags in the power-sensitive data tagging system. Indicates by label The artificial constant penalty weight set for asset level. Indicates label For the first A unified high-dimensional feature The probability of membership. Indicates weight The following is aimed at the first A unified high-dimensional feature Predict the label of the output The probability distribution value.

[0012] In a preferred embodiment, this application can be further configured as follows: the sensitive data identification results are subjected to content topology consistency loss constraint and privacy dual discriminant constraint processing, and the relevant sensitive data is dynamically desensitized based on the constraint results to obtain power desensitized data, specifically including: Based on the sensitive data identification results, the relevant sensitive data is input into the pre-constructed data desensitization model for nonlinear feature transformation, and the content topology consistency loss constraint is applied to the desensitized generated features obtained by the transformation. The expression for the content topology consistency loss constraint is as follows: (7) in, This represents the topology consistency loss value. These represent the characteristics generated by desensitization. and original features The corresponding feature mapping vector, The modulus length; Privacy and availability discrimination are performed on the de-identified generated features respectively, and a composite loss is calculated by combining the content topology consistency loss constraint results. The de-identified generated features are updated based on the calculation results. The calculation expression for the composite loss is as follows: (8) in, This represents the composite loss value. Indicates random noise Follows prior distribution Under these conditions, privacy discriminator Desensitization generates features The expected value of the logarithmic probability of determining that it contains real sensitive information. This indicates that during the generator optimization process, in the presence of random noise... Follows prior distribution Under these conditions, availability discriminator Desensitization generates features The expected value of the logarithmic probability that the data conforms to the actual business distribution pattern is used to guide the generator to output anonymized data that retains the statistical characteristics of the actual business. Representing original features Desensitization generation characteristics Topology consistency loss value, , These represent the business weight and the topology loss balance coefficient, respectively. Based on the updated desensitized generation features, privacy dual-discrimination constraints are applied to the discrimination loss in the privacy discrimination and availability discrimination processes, respectively. When the discrimination adversarial loss reaches a dynamic equilibrium, the desensitized generation features are output to obtain the desensitized power data. The expression for the privacy dual-discrimination constraint is as follows: (9) (10) in, , These represent privacy discriminators. Loss value and availability discriminator Loss value This indicates adherence to the actual distribution of electricity business. Original input features Find the expected value. This indicates that the pair follows a prior distribution. random noise Find the expected value. This indicates that the privacy discriminator will use the original features The probability of being identified as containing real sensitive information. This indicates that the privacy discriminator will generate de-identified features. The probability of being identified as containing real sensitive information. and These represent the privacy discriminator accurately determining the original features. Contains sensitive information and accurately determines the characteristics generated by desensitization. Log probability excluding sensitive information and They represent and The mathematical expectation in their respective sample distribution spaces The availability discriminator will use the original features The probability of determining whether it conforms to the statistical patterns of actual power business distribution. This indicates that the availability discriminator will generate de-identified features. The probability of determining whether it conforms to the statistical patterns of actual power business distribution. and These represent the accuracy of the availability discriminator in determining the original features. To ensure accurate distribution of data and determine the characteristics generated by desensitization The log probability of non-true distributed data. and They represent and The mathematical expectation in their respective sample distribution spaces.

[0013] In a preferred embodiment, this application can be further configured as follows: performing a quantitative assessment of the desensitization risk of the power desensitized data, and based on the assessment results, performing dynamic adjustment of the desensitization mask and adaptive closed-loop update processing of the power desensitized data to obtain the final power desensitized data, specifically including: Calculate the risk quantification score of the de-identified power data, and conduct a de-identification risk quantification assessment based on the risk quantification score to obtain the de-identification risk assessment result. The calculation expression of the risk quantification score is as follows: (11) in, This indicates a quantitative risk score. This represents the Sigmoid activation function. , , , To quantify the weighting coefficients, The label probability distribution score represents the data used for de-identifying electricity data. Indicates the frequency of access to the logarithmic map. This represents a score indicating the user's operational behavior. This represents the score of historical events with attenuation effects; The calculation expression for the logarithmic mapping access frequency is as follows: (12) in, Indicates access frequency; The calculation expression for the historical event score containing the attenuation effect is as follows: (13) in, Represents the time decay constant. Indicates the first The time interval between the occurrence of each historical security incident and the current system risk assessment. Indicates the first The initial harm impact score of the previous historical security incident. Indicates the first The time decay factor corresponding to each historical security event.

[0014] When the desensitization risk assessment result reaches the preset risk threshold, the desensitization mask ratio is adjusted to perform adaptive closed-loop update processing on the power desensitized data to obtain the final power desensitized data.

[0015] Secondly, the above-mentioned inventive objective of this application is achieved through the following technical solutions: A semantic-level sensitive data identification and dynamic desensitization system for large-scale power models, the system comprising: The feature extraction module is used to acquire multi-source heterogeneous power data of the power system and perform cross-modal feature parallel extraction. The extracted multimodal sensitive features are mapped to a high-order tensor product space for implicit semantic association fusion processing to obtain a unified high-dimensional feature set. The sensitive data identification module is used to map the unified high-dimensional feature set one by one into a continuous label distribution vector under the constraint of the conditional mutual information prior matrix and calculate the membership degree of the vector sensitive label to identify sensitive data and obtain the sensitive data identification result. The data desensitization processing module is used to perform content topology consistency loss constraint and privacy dual discrimination constraint processing on the sensitive data identification results respectively, and to perform dynamic desensitization processing on the relevant sensitive data based on the constraint results to obtain power desensitized data; The desensitization quantification assessment module is used to perform a desensitization risk quantification assessment on the power desensitized data, and based on the assessment results, to perform dynamic adjustment of the desensitization mask and adaptive closed-loop update processing of the power desensitized data to obtain the final power desensitized data.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This application achieves cross-modal deep perception and low-level feature fusion of multi-source heterogeneous power data from the power system by performing multimodal fusion, breaking through the limitations of traditional single-modal processing. It realizes cross-modal deep perception and low-level feature fusion of structured and unstructured data, and performs semantic association fusion by mapping features to semantic probability space. It realizes the association capture of sensitive features in complex business scenarios, identifies sensitive data by mapping the continuous label distribution vector of sensitive features and calculating the membership probability distribution, and performs dynamic desensitization by content topology consistency loss constraint and privacy dual discrimination constraint. This allows the desensitized data to retain the statistical features and semantic logic of the original business, improving the business usability of the desensitized data. 2. By quantitatively assessing the risk of desensitization of power desensitized data, the desensitization mask is dynamically adjusted and updated in an adaptive closed loop. An adaptive loop of risk assessment, protection triggering, and strategy updating is constructed to achieve closed-loop management throughout the entire life cycle, thus eliminating the lag of manual intervention in protection. 3. This application uses the LDL paradigm based on CMI constraints to handle scenarios with strong tag coupling and semantic ambiguity. It can deeply mine the implicit semantics of text, effectively reduce the probability of false detection and false negative detection in complex business scenarios, improve the accuracy of semantic recognition, and ensure the safe and stable operation of the power grid while ensuring the security of power data. It is applicable to the protection needs of different sensitive data types, fully adapts to the original architecture of the power system, does not require the deployment of other systems, reduces adaptation and maintenance costs, and has high scalability. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale.

[0018] Figure 1 This is a flowchart illustrating the implementation of the semantic-level sensitive data identification and dynamic desensitization method for large-scale power models in this embodiment.

[0019] Figure 2 This is a block diagram of the semantic-level sensitive data identification and dynamic desensitization system for large power models in this embodiment. Detailed Implementation

[0020] 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, not all, of the embodiments of the present invention. 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.

[0021] It should be understood that, when used in this specification, the terms “comprising” and “including” indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0022] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” are intended to include the plural forms.

[0023] It should also be further understood that the term "and / or" as used in this specification refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes such combinations.

[0024] In one embodiment, such as Figure 1 As shown, this application discloses a semantic-level sensitive data identification and dynamic desensitization method for large-scale power models, which specifically includes the following steps: S1: Acquire multi-source heterogeneous power data from the power system and perform parallel extraction of cross-modal features. Map the extracted multimodal sensitive features to a high-order tensor product space for implicit semantic association fusion processing to obtain a unified high-dimensional feature set.

[0025] Specifically, step S1 includes: S11: Standardize and preprocess the structured data in the multi-source heterogeneous power data to obtain structured features, and extract unstructured features in parallel using multimodal features from the unstructured data in the multi-source heterogeneous power data to obtain unstructured features.

[0026] Specifically, this involves acquiring multi-source heterogeneous power data from the power system, including but not limited to structured and unstructured data. This includes electricity billing data containing user IDs, electricity consumption, and anomaly tags; and unstructured data such as power inspection documents and reports containing highly sensitive business information like relay protection settings, main transformer operating parameters, and dispatch instructions, as well as on-site photos of transformer area inspection equipment. This data is then processed using formulas. Structured data is normalized and converted into standard feature vectors. , These represent the original structured data and the normalized, standardized feature vectors, respectively. , ...

[0027] S12: Map structured and unstructured features to a pre-constructed high-order tensor product space for feature cross-attention fusion processing to construct a feature fusion matrix.

[0028] Specifically, structured and unstructured features are mapped to a pre-constructed high-order tensor product space. Implicit semantic relationships between structured and unstructured features are analyzed through cross-attention fusion, i.e., feature cross-attention fusion operations are performed to construct a feature fusion matrix. The expression for the feature fusion matrix is ​​shown below: (1) in, Represents the feature fusion matrix. , These represent the query projection matrix and the key projection matrix of the cross-attention mechanism, respectively. Indicates the cross-attention head dimension. , These represent structured features and unstructured features, respectively.

[0029] S13: Based on the feature fusion matrix, tensor multiplication is performed on structured and unstructured features to obtain a unified high-dimensional feature set.

[0030] Specifically, the tensor product operation expressions for structured features and unstructured features are as follows: (2) in, Represents unified high-dimensional features. The value projection matrix represents the cross-attention mechanism. Represents the feature fusion matrix. This represents the tensor product operation.

[0031] S2: Under the constraint of the conditional mutual information prior matrix, the unified high-dimensional feature set is mapped one by one to a continuous label distribution vector, and the membership degree of the vector sensitive label is calculated to identify sensitive data and obtain the sensitive data identification result.

[0032] Specifically, in this embodiment, a semantic recognition system based on conditional mutual information optimization of label distribution learning is set up for sensitive data identification. Step S2 includes: S21: Under the constraint of label probability distribution, the unified high-dimensional feature set is mapped one by one to a continuous label distribution vector. The expression of the continuous label distribution vector is as follows: (3) in, Indicates the first A continuous label distribution vector of uniform high-dimensional features from each input. to They represent the first Tag For the first A unified high-dimensional feature The probability of membership. This represents the first in the power sensitive data tagging system. In this embodiment, the power sensitive data tagging system is set to 23 items. =23.

[0033] In this embodiment, the specific calculation process for the probability membership degree is as follows: The first... A unified high-dimensional feature The pre-constructed label distribution mapping fully connected network is input, and after linear transformation and nonlinear activation, it outputs an initial label score vector of dimension m. The initial label score vector is then normalized using the Softmax activation function, mapping the original scores to standardized probability values, thus obtaining the probability membership degree of each sensitive label corresponding to that feature. The normalized probability membership degree naturally satisfies the label probability distribution constraint, i.e., the sum of the probability membership degrees of all labels is 1, and the probability membership degree of a single label ranges from [0,1]. During the model training phase, this probability membership degree will be further optimized in reverse iteratively by combining the conditional mutual information prior matrix constraint and asymmetric divergence loss to strengthen the semantic association constraints between labels and improve the semantic accuracy of membership degree calculation.

[0034] The expression for the label probability distribution constraint is as follows: (4) S22: Apply conditional mutual information prior matrix constraints to the continuous label distribution vector and calculate the asymmetric divergence loss value of the predicted label distribution during the feature label mapping process to obtain the comprehensive loss coefficient.

[0035] Specifically, the expression for the conditional mutual information prior matrix constraint is as follows: (5) in, Indicates conditional mutual information. Represents environment variables for business scenarios. This refers to specific environment variables within the business scenario environment variables. Representing environment variables The prior probability of occurrence, These represent environment variables. Next tag ,Label The conditional probability of occurrence Representing environment variables Next tag ,Label The joint probability of co-occurrence, in this embodiment, is obtained by performing global word frequency statistics and co-occurrence matrix statistics on historical data.

[0036] In this embodiment, the conditional mutual information prior matrix is ​​used as a regularization loss term in the transformation of prior knowledge into a large power model. During backpropagation, it provides higher attention weights for label association pairs with higher conditional mutual information values, thereby suppressing interference from redundant labels.

[0037] The expression for calculating the asymmetric divergence loss is as follows: (6) in, Represents the asymmetric divergence value. This represents the number of features in a unified high-dimensional feature set. This indicates the number of tags in the power-sensitive data tagging system. Indicates by label The artificial constant penalty weight set for asset level. Indicates label For the first A unified high-dimensional feature The probability of membership. Indicates weight The following is aimed at the first A unified high-dimensional feature Predict the label of the output The probability distribution value.

[0038] S23: Based on the comprehensive loss coefficient, the mapping parameters of the continuous label distribution vector are optimized in reverse and the membership degree of sensitive labels is calculated. When the membership degree of a sensitive label exceeds the preset sensitivity judgment threshold, it is marked as sensitive data, and the sensitive data identification result is obtained.

[0039] Specifically, the weighted sum of the asymmetric divergence loss and the conditional mutual information regularization term is used as the comprehensive loss coefficient, and the label distribution is mapped to the weight parameters of the network through the backpropagation algorithm. Iterative optimization is performed. After each round of optimization, the continuous label distribution vector output by the current network, corrected by conditional mutual information prior constraints, is extracted. The probability membership degrees corresponding to all power-sensitive labels in the vector are extracted and weighted according to the sensitive asset level weights of each label to obtain the membership degree of the sensitive label corresponding to the feature. The membership degree of the sensitive label is compared with a preset sensitivity judgment threshold. When it exceeds the preset threshold, the corresponding data is judged as sensitive data. In this embodiment, the preset sensitivity threshold is set to [0.6, 0.8], where 0.6 is the general sensitive data judgment threshold and 0.8 is the core sensitive data judgment threshold. A membership degree exceeding 0.6 is marked as sensitive data, and a membership degree exceeding 0.8 is further marked as high-sensitivity data.

[0040] S3: The sensitive data identification results are subjected to content topology consistency loss constraint and privacy dual discrimination constraint processing respectively. Based on the constraint results, the relevant sensitive data is dynamically desensitized to obtain power desensitized data.

[0041] Furthermore, to avoid losing the core semantics of electricity during the desensitization process, the desensitization-generated features are... and original features Content topology consistency constraints are applied, and features are generated by extracting de-identified data. and original features The activation feature vectors in the high-dimensional manifold space of the power large model are constrained by cosine similarity to ensure that the de-identified data have the same semantic topological features when processed by the large model.

[0042] Specifically, step S3 includes: S31: Based on the sensitive data identification results, the relevant sensitive data is input into the pre-constructed data desensitization model for nonlinear feature transformation, and the content topology consistency loss constraint is applied to the desensitized generated features obtained by the transformation.

[0043] The expression for the content topology consistency loss constraint is as follows: (7) in, This represents the topology consistency loss value. , These represent the characteristics generated by desensitization. and original features The corresponding feature mapping vector, They represent The modulus length; Specifically, this embodiment constructs an adaptive dual discriminator cognitive adversarial network (DDC-GAN) as a data desensitization model for dynamic desensitization. The data desensitization model includes a generative model based on the Transformer architecture, a privacy discriminator based on the CNN architecture, and an availability discriminator based on the self-attention architecture. Specifically, based on the sensitive data identification results, the relevant sensitive data, i.e., the original features and random noise following a standard normal distribution, are input into the generative model based on the Transformer architecture to perform high-dimensional space transformation on the sensitive data, and output the reconstructed non-sensitive feature vector.

[0044] S32: Perform privacy discrimination and availability discrimination on the de-identified generated features respectively, calculate the composite loss by combining the content topology consistency loss constraint results, and update the de-identified generated features based on the calculation results.

[0045] The formula for calculating the composite loss is as follows: (8) in, This represents the composite loss value. Indicates random noise Follows prior distribution Under these conditions, privacy discriminator Desensitization generates features The expected value of the logarithmic probability of determining that it contains real sensitive information; This indicates that during the generator optimization process, in the presence of random noise... Follows prior distribution Under these conditions, availability discriminator Desensitization generates features The expected value of the logarithmic probability that the data conforms to the actual business distribution pattern is used to guide the generator to output anonymized data that retains the statistical characteristics of the actual business. Representing original features Desensitization generation characteristics The topology consistency loss value, , These represent the business weight and the topology loss balance coefficient, respectively. The topology loss balance coefficient is a constant with a value range of 0.5-1.5, used to adjust the trade-off between erasing privacy and retaining business characteristics.

[0046] Specifically, the insensitive feature vector and the corresponding original sensitive feature are input into an adaptive dual discriminator cognitive adversarial network. The privacy discriminator determines the probability that the insensitive feature vector contains a sensitive label, and the availability discriminator determines the probability that the insensitive feature vector conforms to the statistical law of the distribution of real power business. A composite loss is calculated, and the result of the composite loss calculation is fed back to the generative model based on the Transformer architecture for parameter update, thereby updating the output desensitized generated features.

[0047] S33: Based on the updated desensitized generation features, calculate the discrimination loss in the privacy discrimination and availability discrimination processes respectively, and apply privacy dual discrimination constraints. When the discrimination adversarial loss reaches dynamic equilibrium, output the desensitized generation features to obtain the power desensitized data.

[0048] The expression for the privacy dual-discrimination constraint is as follows: (9) (10) in, , These represent privacy discriminators. Loss value and availability discriminator Loss value This indicates adherence to the actual distribution of electricity business. Original input features Find the expected value. This indicates that the pair follows a prior distribution. random noise The difference between the two methods lies in the sample space used for calculation: the former is based on the real original feature dataset, while the latter is based on a generated set of random noise. This indicates that the privacy discriminator will use the original features The probability of being identified as containing real sensitive information. This indicates that the privacy discriminator will generate de-identified features. The probability of being identified as containing real sensitive information. and These represent the privacy discriminator accurately determining the original features. Contains sensitive information and accurately determines the characteristics generated by desensitization. Log probability excluding sensitive information and They represent and The mathematical expectation in their respective sample distribution spaces; The availability discriminator will use the original features The probability of determining whether it conforms to the statistical patterns of actual power business distribution. This indicates that the availability discriminator will generate de-identified features. The probability of determining whether it conforms to the statistical patterns of actual power business distribution. and These represent the accuracy of the availability discriminator in determining the original features. For real-world data distribution and to accurately determine the characteristics of desensitization generation The log probability of non-true distributed data. and They represent and The mathematical expectation under their respective sample distribution spaces. This embodiment adopts an adversarial training mechanism: for the availability discriminator, its training objective is to accurately distinguish between real samples and generated samples, that is, the probability of judging the original real power business feature output approaches 1, and the probability of judging the de-identified generated feature output approaches 0; the optimization objective of the generator is the opposite, aiming to induce the availability discriminator to judge the de-identified generated features as real business distribution data, and the two are trained alternately and iteratively until a dynamic equilibrium is reached.

[0049] S4: Conduct a quantitative assessment of the desensitization risk of the power desensitized data, and based on the assessment results, perform dynamic adjustment of the desensitization mask and adaptive closed-loop update processing of the power desensitized data to obtain the final power desensitized data.

[0050] Specifically, step S4 includes: S41: Calculate the risk quantification score of the de-identified power data, conduct a de-identification risk quantification assessment based on the risk quantification score, and obtain the de-identification risk assessment result. The calculation expression of the risk quantification score is as follows: (11) in, This indicates a quantitative risk score. This represents the Sigmoid activation function. , , , To quantify the weight coefficients, the eigenvectors with the highest probabilities are obtained by constructing a judgment matrix using the analytic hierarchy process (AHP). The sum of all quantized weight coefficients is 1. The label probability distribution score of the de-identified power data is obtained from the maximum label probability of the de-identified power data. Indicates the frequency of access to the logarithmic map. The score represents the user's action behavior, which is obtained by extracting the user's action sequence and calculating the anomaly transition probability using a hidden Markov model. This represents the score of historical events with attenuation effects.

[0051] The expression for calculating the logarithmic map access frequency is as follows: (12) in, Indicates the frequency of access.

[0052] The calculation expression for the historical event score including the attenuation effect is shown below: (13) in, The time decay constant is a preset positive real number used to control the decay rate of the impact of the historical security events on the current risk quantification score over time. Indicates the first The time interval between the occurrence of each historical security incident and the current system risk assessment; Indicates the first The initial harm impact score of each historical security incident is quantified and assigned within a preset score range based on the severity of the historical security incident. Indicates the first The time decay factor corresponding to each historical security event is used to make the contribution weight of the historical security event that occurred a long time ago to the current data breach risk quantification score decrease exponentially.

[0053] S42: When the desensitization risk assessment result reaches the preset risk threshold, the desensitization mask ratio is adjusted to perform adaptive closed-loop update processing on the power desensitization data to obtain the final power desensitization data.

[0054] Specifically, this embodiment employs a tiered protection mechanism: risk quantification scoring is used to classify risk levels and match corresponding desensitization strengths, with a preset risk threshold serving as the critical value for triggering adaptive closed-loop updates. The scores for each risk dimension are first centrally standardized and then weighted and summed, then mapped to the [0,1] interval using a Sigmoid activation function to obtain the final risk quantification score. The closer the value is to 1, the higher the risk of data leakage. The preset risk threshold is set to 0.9. When the risk quantification score is 0.27, it is considered low risk, and only a lightweight partial mask is performed; when the risk quantification score is 0.84, it is considered high risk, triggering full feature reconstruction and desensitization processing; when the risk quantification score reaches the preset threshold of 0.9, a deterministic data leakage risk is identified, and the desensitization mask ratio is further adjusted based on full desensitization, performing adaptive closed-loop updates on the power desensitized data. Furthermore, this embodiment also sets a risk increment threshold. When the risk increment exceeds the risk increment threshold, a closed-loop update strategy is also triggered to adaptively update the power desensitized data, achieving self-purification of the protection strategy and obtaining the final power desensitized data.

[0055] In this embodiment, the risk threshold is obtained by statistically analyzing the risk values ​​under normal business conditions over the past three months using a normal distribution. The upper limit of the 99.7% confidence interval is taken as the risk threshold to determine whether there is a risk of deterministic data leakage. The average value obtained by taking the derivative of the risk increment slope of the power system five minutes before the occurrence of historical known network attacks or abnormal crawling events is taken as the risk increment threshold to prevent automated high-frequency data theft before the absolute risk value reaches the upper limit.

[0056] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0057] In one embodiment, a semantic-level sensitive data identification and dynamic de-identification system for large-scale power models is provided. This system corresponds one-to-one with the semantic-level sensitive data identification and dynamic de-identification method for large-scale power models described in the above embodiments. Figure 2 As shown, this semantic-level sensitive data identification and dynamic desensitization system for large-scale power models includes a feature extraction module, a sensitive data identification module, a data desensitization processing module, and a desensitization quantification evaluation module. Detailed descriptions of each functional module are as follows: The feature extraction module is used to acquire multi-source heterogeneous power data from the power system and perform parallel cross-modal feature extraction. The extracted multimodal sensitive features are mapped to a high-order tensor product space for implicit semantic association fusion processing to obtain a unified high-dimensional feature set.

[0058] The sensitive data identification module is used to map a unified high-dimensional feature set into continuous label distribution vectors one by one under the constraint of a conditional mutual information prior matrix, and to calculate the membership degree of the vector sensitive labels to identify sensitive data and obtain the sensitive data identification results.

[0059] The data desensitization module is used to perform content topology consistency loss constraint and privacy dual discrimination constraint processing on the sensitive data identification results, and to perform dynamic desensitization processing on the relevant sensitive data based on the constraint results to obtain power desensitized data.

[0060] The desensitization quantification assessment module is used to quantitatively assess the desensitization risk of power desensitized data. Based on the assessment results, the desensitization mask is dynamically adjusted and the power desensitized data is updated adaptively in a closed loop to obtain the final power desensitized data.

[0061] Specific limitations regarding the semantic-level sensitive data identification and dynamic desensitization system for large-scale power model systems can be found in the limitations of the semantic-level sensitive data identification and dynamic desensitization method for large-scale power model systems described above, and will not be repeated here. Each module in the aforementioned semantic-level sensitive data identification and dynamic desensitization system for large-scale power model systems can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0062] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application of the technical solution and the constraints involved. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0063] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0064] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0065] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for semantic-level sensitive data identification and dynamic desensitization in large-scale power models, characterized in that, The method includes: Multi-source heterogeneous power data from the power system is acquired and cross-modal features are extracted in parallel. The extracted multi-modal sensitive features are mapped to a high-order tensor product space for implicit semantic association fusion processing to obtain a unified high-dimensional feature set. Under the constraint of the conditional mutual information prior matrix, the unified high-dimensional feature set is mapped one by one to a continuous label distribution vector and the membership degree of the vector sensitive label is calculated to identify sensitive data and obtain the sensitive data identification result. The sensitive data identification results are subjected to content topology consistency loss constraint and privacy dual discrimination constraint processing respectively. Based on the constraint results, the relevant sensitive data are dynamically desensitized to obtain power desensitized data. The power desensitized data is subjected to a quantitative assessment of desensitization risk. Based on the assessment results, the desensitization mask is dynamically adjusted and the power desensitized data is updated adaptively in a closed loop to obtain the final power desensitized data.

2. The semantic-level sensitive data identification and dynamic desensitization method for large-scale power models according to claim 1, characterized in that, The process involves acquiring multi-source heterogeneous power data from the power system and performing parallel cross-modal feature extraction. The extracted multimodal sensitive features are then mapped to a high-order tensor product space for implicit semantic association fusion processing, resulting in a unified high-dimensional feature set. Specifically, this includes: The structured data in the multi-source heterogeneous power data is standardized and preprocessed to obtain structured features, and the unstructured data in the multi-source heterogeneous power data is extracted in parallel using multimodal features to obtain unstructured features. The structured and unstructured features are mapped to a pre-constructed high-order tensor product space for feature cross-attention fusion processing to construct a feature fusion matrix; Based on the feature fusion matrix, tensor multiplication is performed on the structured features and the unstructured features to obtain a unified high-dimensional feature set.

3. The semantic-level sensitive data identification and dynamic desensitization method for large-scale power models according to claim 2, characterized in that, The step of mapping the structured and unstructured features to a pre-constructed high-order tensor product space for feature cross-attention fusion processing to construct a feature fusion matrix specifically includes: The expression for the feature fusion matrix is ​​as follows: (1) in, Represents the feature fusion matrix. , These represent the query projection matrix and the key projection matrix for the cross-attention mechanism, respectively. Indicates the cross-attention head dimension. , These represent structured features and unstructured features, respectively.

4. The semantic-level sensitive data identification and dynamic desensitization method for large-scale power models according to claim 3, characterized in that, The step of performing tensor multiplication on the structured features and the unstructured features based on the feature fusion matrix to obtain a unified high-dimensional feature set specifically includes: The tensor product expression for the structured features and the unstructured features is as follows: (2) in, Represents unified high-dimensional features. The value projection matrix represents the cross-attention mechanism. Represents the feature fusion matrix. This represents the tensor product operation.

5. The semantic-level sensitive data identification and dynamic desensitization method for large-scale power models according to claim 1, characterized in that, Under the constraint of the conditional mutual information prior matrix, the unified high-dimensional feature set is mapped one by one to a continuous label distribution vector, and the membership degree of the vector sensitive label is calculated to identify sensitive data, thereby obtaining the sensitive data identification result. Specifically, this includes: Under the constraint of label probability distribution, the unified high-dimensional feature set is mapped one by one to a continuous label distribution vector, and the expression of the continuous label distribution vector is as follows: (3) in, Indicates the first A continuous label distribution vector of uniform high-dimensional features from each input. to They represent the first Tag For the first A unified high-dimensional feature The probability of membership. This represents the first in the power sensitive data tagging system. A specific label; The expression for the tag probability distribution constraint is as follows: (4) The continuous label distribution vector is constrained by a conditional mutual information prior matrix, and the asymmetric divergence loss value of the predicted label distribution is calculated during the feature label mapping process to obtain the comprehensive loss coefficient. Based on the comprehensive loss coefficient, the mapping parameters of the continuous label distribution vector are back-optimized and the membership degree of sensitive labels is calculated. When the membership degree of the sensitive label exceeds the preset sensitivity judgment threshold, it is marked as sensitive data, and the sensitive data identification result is obtained.

6. The semantic-level sensitive data identification and dynamic desensitization method for large-scale power models according to claim 5, characterized in that, The process of applying conditional mutual information prior matrix constraints to the continuous label distribution vector and calculating the asymmetric divergence loss value of the predicted label distribution during feature label mapping to obtain the conditional mutual information prior matrix constraints in the comprehensive loss coefficient specifically includes: The expression for the conditional mutual information prior matrix constraint is as follows: (5) in, Indicates conditional mutual information. Represents environment variables for business scenarios. This refers to specific environment variables within the business scenario environment variables. Representing environment variables The prior probability of occurrence, These represent environment variables. Next tag ,Label The conditional probability of occurrence Representing environment variables Next tag ,Label Joint probability of co-occurrence.

7. The semantic-level sensitive data identification and dynamic desensitization method for large-scale power models according to claim 5, characterized in that, The process of constraining the continuous label distribution vector with a conditional mutual information prior matrix and calculating the asymmetric divergence loss value of the predicted label distribution during feature label mapping to obtain the asymmetric divergence loss value in the comprehensive loss coefficient specifically includes: The expression for calculating the asymmetric divergence loss value is as follows: (6) in, Represents the asymmetric divergence value. This represents the number of features in a unified high-dimensional feature set. This indicates the number of tags in the power-sensitive data tagging system. Indicates by label The artificial constant penalty weight set for asset level. Indicates label For the first A unified high-dimensional feature The probability of membership. Indicates in weight The following is aimed at the first A unified high-dimensional feature Predict the label of the output The probability distribution value.

8. The semantic-level sensitive data identification and dynamic desensitization method for large-scale power models according to claim 5, characterized in that, The sensitive data identification results are subjected to content topology consistency loss constraint and privacy dual discriminant constraint processing, respectively. Based on the constraint results, the relevant sensitive data is dynamically de-identified to obtain power de-identified data, specifically including: Based on the sensitive data identification results, the relevant sensitive data is input into the pre-constructed data desensitization model for nonlinear feature transformation, and the content topology consistency loss constraint is applied to the desensitized generated features obtained by the transformation. The expression for the content topology consistency loss constraint is as follows: (7) in, This represents the topology consistency loss value. , These represent the characteristics generated by desensitization. and original features The corresponding feature mapping vector, They represent The modulus length; Privacy and availability discrimination are performed on the de-identified generated features respectively, and composite loss is calculated by combining the content topology consistency loss constraint results. The de-identified generated features are updated based on the calculation results. The calculation expression for the composite loss is as follows: (8) in, This represents the composite loss value. Indicates random noise Follows prior distribution Under these conditions, privacy discriminator Desensitization generates features The expected value of the logarithmic probability of determining that it contains real sensitive information. This indicates that during the generator optimization process, in the presence of random noise... Follows prior distribution Under these conditions, availability discriminator Desensitization generates features The expected value of the logarithmic probability that is determined to conform to the actual business distribution pattern. Representing original features Desensitization generation characteristics The topology consistency loss value, , These represent the business weight and the topology loss balance coefficient, respectively. Based on the updated desensitized generation features, privacy dual-discrimination constraints are applied to the discrimination loss in the privacy discrimination and availability discrimination processes, respectively. When the discrimination adversarial loss reaches a dynamic equilibrium, the desensitized generation features are output to obtain the desensitized power data. The expression for the privacy dual-discrimination constraint is as follows: (9) (10) in, , These represent privacy discriminators. Loss value and availability discriminator Loss value This indicates adherence to the actual distribution of electricity business. Original input features Find the expected value. This indicates that the pair follows a prior distribution. random noise Find the expected value. This indicates that the privacy discriminator will use the original features The probability of being identified as containing real sensitive information. This indicates that the privacy discriminator will generate de-identified features. The probability of being identified as containing real sensitive information. and These represent the privacy discriminator accurately determining the original features. Contains sensitive information and accurately determines the characteristics generated by desensitization. Log probability excluding sensitive information and They represent and The mathematical expectation in their respective sample distribution spaces The availability discriminator will use the original features The probability of determining whether it conforms to the statistical patterns of actual power business distribution. This indicates that the availability discriminator will generate de-identified features. The probability of determining whether it conforms to the statistical patterns of actual power business distribution. and These represent the accuracy of the availability discriminator in determining the original features. To ensure accurate distribution of data and determine the characteristics generated by desensitization The log probability of non-true distribution data. and They represent and The mathematical expectation in their respective sample distribution spaces.

9. The semantic-level sensitive data identification and dynamic desensitization method for large-scale power models according to claim 1, characterized in that, The process of performing a quantitative assessment of the desensitization risk on the desensitized power data, and then dynamically adjusting the desensitization mask and performing adaptive closed-loop update processing on the desensitized power data based on the assessment results to obtain the final desensitized power data, specifically includes: Calculate the risk quantification score of the de-identified power data, and conduct a de-identification risk quantification assessment based on the risk quantification score to obtain the de-identification risk assessment result. The calculation expression of the risk quantification score is as follows: (11) in, This indicates a quantitative risk score. This represents the Sigmoid activation function. , , , To quantify the weighting coefficients, The label probability distribution score represents the data used for de-identifying electricity data. Indicates the frequency of access to the logarithmic map. This represents a score indicating the user's operational behavior. This represents the score of historical events with attenuation effects; The calculation expression for the logarithmic mapping access frequency is as follows: (12) in, Indicates access frequency; The calculation expression for the historical event score containing the attenuation effect is as follows: (13) in, Represents the time decay constant. Indicates the first The time interval between the occurrence of each historical security incident and the current system risk assessment. Indicates the first The initial harm impact score of the previous historical security incident. Indicates the first The time decay factor corresponding to each historical security event. When the desensitization risk assessment result reaches the preset risk threshold, the desensitization mask ratio is adjusted to perform adaptive closed-loop update processing on the power desensitized data to obtain the final power desensitized data.

10. A semantic-level sensitive data identification and dynamic desensitization system for large-scale power models, characterized in that, The system includes: The feature extraction module is used to acquire multi-source heterogeneous power data of the power system and perform cross-modal feature parallel extraction. The extracted multimodal sensitive features are mapped to a high-order tensor product space for implicit semantic association fusion processing to obtain a unified high-dimensional feature set. The sensitive data identification module is used to map the unified high-dimensional feature set one by one into a continuous label distribution vector under the constraint of the conditional mutual information prior matrix and calculate the membership degree of the vector sensitive label to identify sensitive data and obtain the sensitive data identification result. The data desensitization processing module is used to perform content topology consistency loss constraint and privacy dual discrimination constraint processing on the sensitive data identification results respectively, and to perform dynamic desensitization processing on the relevant sensitive data based on the constraint results to obtain power desensitized data; The desensitization quantification assessment module is used to perform a desensitization risk quantification assessment on the power desensitized data, and based on the assessment results, to perform dynamic adjustment of the desensitization mask and adaptive closed-loop update processing of the power desensitized data to obtain the final power desensitized data.