A data security protection method and system

Abstract

The present application relates to the technical field of data protection, in particular to a data security protection method and system, which are applied to data security protection of public opinion monitoring and analysis model. First, the original public opinion data is cleaned to eliminate invalid information, and core entities, associated entities and hierarchical regional entities are extracted to construct an entity feature label library. Then, based on the entity feature label library, the total value of entity sensitivity and the total value of public opinion risk are calculated, and combined with the public opinion transmission stage, the public opinion risk level is calculated and classified. Next, according to the determined public opinion risk level and the corresponding transmission stage, differential desensitization processing is carried out. Finally, the desensitized public opinion data is input into the public opinion monitoring and analysis model to realize effective protection of sensitive entity information, avoid leakage or reverse tracing of sensitive information during model processing, and realize data security protection of the public opinion monitoring and analysis model.

Description

Technical Field

[0001] This invention relates to the field of data protection technology, specifically to a data security protection method and system. Background Technology

[0002] Currently, public opinion monitoring and analysis models built on artificial intelligence platforms can achieve functions such as public opinion early warning, topic identification, and subject / regional analysis. They can complete statistical analysis of public opinion data, hotspot mining, and invalid data filtering, and conduct comprehensive situational assessment of public opinion events, including initial information, dissemination channels, and development trends. Simultaneously, they possess real-time early warning capabilities, learning and training from historical public opinion work order data to identify sensitive information and issue timely warnings. The massive amounts of public opinion data processed by these models contain various sensitive information, including core entities and regional entities. During processing, these models are susceptible to security risks such as sensitive information leakage and reverse engineering. Therefore, providing professional and precise data security protection for public opinion monitoring and analysis models is of great significance.

[0003] Commonly used existing data security protection methods include static anonymization, federated learning, and differential privacy methods. However, traditional static anonymization is prone to destroying the core analytical features of public opinion, such as category, administrative level, and geographical level, resulting in a significant decrease in the accuracy of model identification and early warning. General privacy computing schemes such as federated learning and homomorphic encryption have high computational overhead and high inference latency, which cannot meet the real-time requirements of public opinion monitoring. They are also prone to insufficient protection of sensitive entities in the model, and can easily lead to information leakage by reverse inferring core sensitive entities through related information. Summary of the Invention

[0004] The purpose of this invention is to provide a data security protection method and system for public opinion monitoring and analysis models.

[0005] The technical solution of this invention is as follows:

[0006] A data security protection method applied to the data security protection of a public opinion monitoring and analysis model includes the following operations:

[0007] S1. Obtain raw public opinion data, remove invalid data, extract core entities, related entities, and regional entities to form an entity feature tag library; the regional entity levels include provincial regional entities, prefecture-level regional entities, and district / county-level regional entities.

[0008] S2. Based on the entity feature tag library, obtain the subject sensitivity value and the region sensitivity value; merge the subject sensitivity value and the region sensitivity value to obtain the total entity sensitivity value; obtain the basic public opinion risk value and the public opinion dissemination stage risk value, and after weighted fusion, obtain the total public opinion risk value; the public opinion dissemination stage includes: budding, fermentation, outbreak, and decline; match the correction coefficient according to the public opinion dissemination stage, and based on the correction coefficient, merge the total entity sensitivity value and the total public opinion risk value to obtain the comprehensive sensitivity risk value, and obtain the public opinion risk level through threshold classification;

[0009] S3. Based on the public opinion risk level and the corresponding stage of public opinion dissemination, the core entities and related entities are anonymized to obtain anonymized public opinion data. Anonymization includes irreversible vector mapping replacement, and / or local feature fuzzing, and / or feature masking. The anonymized public opinion data is input into the public opinion monitoring and analysis model to obtain the anonymized public opinion analysis results.

[0010] In S3, when the public opinion risk level is high and the public opinion dissemination stage is in its nascent or receding phase, the core entity is replaced with a high-intensity irreversible vector mapping, and the related entities are subject to lightweight local feature blurring. When the public opinion risk level is high and the public opinion dissemination stage is in its fermentation or outbreak phase, the core entity is replaced with an irreversible vector mapping and its features are masked, and the related entities are replaced with an irreversible vector mapping. When the public opinion risk level is medium and the public opinion dissemination stage is in its nascent or receding phase, the core entity and the related entities are respectively subject to local feature blurring, while retaining the regional administrative level. When the public opinion risk level is medium and the public opinion dissemination stage is in its fermentation or outbreak phase, the core entity is replaced with an irreversible vector mapping, and the related entities are subject to local feature blurring. When the public opinion risk level is low and the public opinion dissemination stage is in its nascent or receding phase, the original expression is retained. When the public opinion risk level is low and the public opinion dissemination stage is in its fermentation or outbreak phase, the core entity and the related entities are respectively subject to local feature blurring.

[0011] In S3, the irreversible vector mapping replacement operation is as follows: using a preset irreversible hash vector mapping algorithm, the entity's identity representation information is mapped to a fixed-dimensional anonymous vector; the local feature fuzzification operation is as follows: retaining the entity's preset necessary business features, and performing fuzzy replacement processing on the entity's preset local non-essential sensitive features; the feature masking operation is as follows: removing the entity's core features, including sensitive attributes, key identity features, and associated semantic information.

[0012] The method for obtaining the subject sensitivity value in S2 is as follows: Based on effective public opinion data, construct an entity relationship graph; in the entity relationship graph, the nodes are core entities and related entities, and the edge weights are the relationship strength between entities; based on the paths between core entity nodes and related entity nodes in the entity relationship graph, obtain the risk diffusion coefficient; multiply the risk diffusion coefficient by the initial subject sensitivity value of the core entity to obtain the subject sensitivity value.

[0013] In S2, the method for obtaining the regional sensitivity value is as follows: A regional prior sensitivity database is established, and the prior sensitivity value of regional entities is calculated; within a preset time window, the multi-source attention of regional entities is statistically analyzed, and for each source, the historical mean and standard deviation of attention are statistically analyzed, and the logarithmic relative anomaly is calculated. The mean of the logarithmic relative anomalies from all sources is taken as the comprehensive anomaly; a regional relationship graph is constructed, where nodes are regional entities, and edge weights are obtained based on node geographic adjacency and cross-platform propagation coupling; the comprehensive anomaly is diffused based on the edge weights to obtain the spatial spillover; the child regional attention distribution of the regional entity that serves as the parent regional entity is obtained, and the hierarchical clustering degree is obtained based on the attention distribution of the child regional entities; the prior sensitivity value, comprehensive anomaly, hierarchical clustering degree, and spatial spillover are weighted and fused to obtain the regional sensitivity value.

[0014] In S3, after obtaining the de-identified public opinion data, it also includes obtaining the semantic similarity between the original public opinion data and the de-identified public opinion data. If the semantic similarity is not less than the semantic similarity threshold, the verification passes; if the semantic similarity is less than the semantic similarity threshold, for entities with blurred local features, the full name blurring is changed to only individual identifier blurring, while retaining industry, level, and function; for feature-masked fragments, only key identity information is masked.

[0015] The core entity is the dominant party in the public opinion event within the valid public opinion data; related entities refer to entities that have a direct relationship with the core entity.

[0016] A data security protection system, deployed within a public opinion monitoring and analysis model, is used to implement the aforementioned data security protection method, comprising:

[0017] The entity feature tag library generation module is used to acquire raw public opinion data, remove invalid data, and extract core entities, related entities, and regional entities to form an entity feature tag library; the regional entity levels include provincial regional entities, prefecture-level regional entities, and district / county-level regional entities.

[0018] The public opinion risk level generation module is used to obtain subject sensitivity values ​​and regional sensitivity values ​​based on an entity feature tag library; merge the subject sensitivity values ​​and regional sensitivity values ​​to obtain the total entity sensitivity value; obtain the basic public opinion risk value and the risk value of the public opinion dissemination stage, and obtain the total public opinion risk value through weighted fusion; the public opinion dissemination stages include: budding, fermentation, outbreak, and decline; match the correction coefficient according to the public opinion dissemination stage, and merge the total entity sensitivity value and the total public opinion risk value based on the correction coefficient to obtain the comprehensive sensitivity risk value, and obtain the public opinion risk level through threshold classification;

[0019] The desensitized public opinion analysis result generation module is used to desensitize core entities and related entities according to the public opinion risk level and the corresponding public opinion dissemination stage to obtain desensitized public opinion data. The desensitization process includes irreversible vector mapping replacement, and / or local feature blurring, and / or feature masking. The desensitized public opinion data is input into the public opinion monitoring and analysis model to obtain the desensitized public opinion analysis results.

[0020] A data security protection device includes a processor and a memory, wherein the processor implements the above-described data security protection method when executing a computer program stored in the memory.

[0021] A computer-readable storage medium for storing a computer program, wherein the computer program implements the above-described data security protection method when executed by a processor.

[0022] The beneficial effects of this invention are as follows:

[0023] This invention provides a data security protection method applied to the data security protection of a public opinion monitoring and analysis model. First, the original public opinion data is cleaned to remove invalid information, and core entities, related entities, and hierarchical regional entities are extracted to construct an entity feature tag library, laying a standardized and structured foundation for the data security protection of the public opinion monitoring and analysis model. Then, based on this tag library, the total sensitivity value of entities and the total public opinion risk value are quantified. Combined with the public opinion propagation stage, the public opinion risk level is calculated and graded, providing a reliable basis for subsequent precise and differentiated security desensitization. Next, according to the determined public opinion risk level and corresponding propagation stage, irreversible vector mapping replacement, or local feature fuzzification, or feature masking are used for core entities and related entities to achieve differentiated desensitization processing, blocking the leakage of sensitive information and the risk of reverse tracing from the source. Finally, the desensitized public opinion data is input into the public opinion monitoring and analysis model, enabling the model to effectively protect sensitive information during the processing of public opinion data, preventing the leakage or reverse tracing of sensitive identity information, and achieving full-link data security and privacy protection for the public opinion monitoring and analysis model without affecting its normal analytical capabilities.

[0024] This invention provides a data security protection method that includes niche related entities and core entities within a unified protection scope, avoiding the risk of leakage of core sensitive entities by reverse deducing from related information, and achieving full coverage of public opinion entity protection. Simultaneously, in calculating the comprehensive sensitivity risk value used for public opinion risk level classification, the characteristics of entities at the provincial, municipal, and county levels are considered, and regional sensitivity values ​​are calculated, solving the imbalance problem of insufficient protection for niche sensitive entities at the county level and excessive desensitization of general regional entities at the provincial level. Furthermore, the protection intensity is dynamically adjusted according to the propagation stages of public opinion—initiation, fermentation, outbreak, and decline—avoiding the drawbacks of static desensitization affecting handling at the initiation stage and insufficient protection at the outbreak stage, ensuring that security protection is deeply matched with the development trend of public opinion.

[0025] This invention provides a data security protection method that, by constructing a priori regional sensitivity database and quantifying regional sensitivity values, implements differentiated entity desensitization processing based on public opinion risk levels and dissemination stages, and combines this with multi-dimensional semantic consistency verification, can effectively protect sensitive information such as entities and regions from leakage during model processing, prevent the risk of reverse reconstruction of sensitive information, and ensure the semantic consistency of desensitized data. This ensures that data processing is safe and compliant throughout the entire process of entity recognition, regional recognition, issue analysis, and public opinion early warning, providing comprehensive data security support for the stable and secure operation of public opinion monitoring and analysis models. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the exemplary embodiments of this application clearer, the technical solutions in the exemplary embodiments of this application are described clearly and completely below. Obviously, the described exemplary embodiments are only some embodiments of this application, and not all embodiments.

[0027] The method in this embodiment is applied to the data security protection of a public opinion monitoring and analysis model. The public opinion monitoring and analysis model is called the Public Opinion Monitoring Large Model, which has functions such as subject identification, region identification, topic identification, category judgment, event situation analysis, and public opinion early warning. The method in this embodiment includes the following operations:

[0028] S1. Obtain raw public opinion data, remove invalid data, extract core entities, related entities, and regional entities to form an entity feature tag library; the regional entity levels include provincial regional entities, prefecture-level regional entities, and district / county-level regional entities.

[0029] S2. Based on the entity feature tag library, obtain the subject sensitivity value and the region sensitivity value; merge the subject sensitivity value and the region sensitivity value to obtain the total entity sensitivity value; obtain the basic public opinion risk value and the public opinion dissemination stage risk value, and after weighted fusion, obtain the total public opinion risk value; the public opinion dissemination stage includes: budding, fermentation, outbreak, and decline; match the correction coefficient according to the public opinion dissemination stage, and based on the correction coefficient, merge the total entity sensitivity value and the total public opinion risk value to obtain the comprehensive sensitivity risk value, and obtain the public opinion risk level through threshold classification;

[0030] S3. Based on the public opinion risk level and the corresponding stage of public opinion dissemination, the core entities and related entities are desensitized to obtain desensitized public opinion data. Desensitization includes irreversible vector mapping replacement, and / or local feature blurring, and / or feature masking. The desensitized public opinion data is input into the public opinion monitoring and analysis model to obtain the desensitized public opinion analysis results.

[0031] The specific steps are detailed below.

[0032] S1. Obtain raw public opinion data, remove invalid data, extract core entities, related entities, and regional entities to form an entity feature tag library; the regional entity levels include provincial regional entities, prefecture-level regional entities, and district / county-level regional entities.

[0033] First, we acquire raw public opinion data from multiple sources, including social media platforms, news media, industry forums, and user comments. This raw data includes basic information such as title, body text, publication location, publication time, and data source. We then intelligently filter the raw data, removing invalid data such as advertisements, entertainment news, and irrelevant lifestyle sharing. Only valid public opinion data related to the company's operations, business services, and industry development is retained. We focus on protecting and processing the valid public opinion data to improve overall data processing efficiency.

[0034] Then, to achieve full-dimensional extraction of public opinion entities and avoid reverse deduction leakage vulnerabilities caused by protecting only core entities and ignoring niche related entities, this embodiment performs entity semantic analysis on the filtered valid public opinion data, extracts core entities and niche related entities from the data, and labels the extracted core entities and related entities with corresponding entity types, including three categories: personal entities, enterprise entities, and government entities. Core entities refer to the core objects and the leading parties of public opinion events in the valid public opinion data; related entities refer to entities in the valid public opinion data that have direct relationships with core entities, such as business affiliation, regional association, cooperation and support, and scenario dependence, and exist dependent on core entities.

[0035] Meanwhile, to achieve administrative hierarchy and refined extraction of regional entities, breaking away from the traditional, unhierarchical regional entity identification model, this embodiment performs regional semantic analysis on valid public opinion data to extract regional entities. These entities are then categorized into three levels according to administrative hierarchy: provincial, municipal, and district / county level. Furthermore, the extracted regional entities are labeled with high, medium, and low levels of public opinion attention. For example, if the valid public opinion data is "XX Province Power Company launches a new smart electricity package, simultaneously promoted throughout the province," then the extracted regional entity is XX Province, classified as a provincial-level entity, and labeled with a medium level of public opinion attention.

[0036] Finally, the core entities, related entities, and regional entities extracted and annotated above are structurally integrated and annotated, along with corresponding annotations including entity name, entity type, regional level, and public opinion attention. Based on standardized annotation rules, an entity feature tag library is constructed, allowing subsequent sensitive risk calculations to directly call the structured data in the tag library.

[0037] This embodiment removes invalid data from the original public opinion data and extracts core entities, related entities, and regional entities to construct a standardized entity feature tag library. It clearly divides the regional entity levels into three levels: provincial, municipal, and county. This can effectively purify the quality of the original data, filter out redundant and interfering information, and achieve the standardization and structured sorting of entity and regional information. This provides accurate and unified basic data support for subsequent entity sensitivity assessment, regional sensitivity calculation, and differentiated desensitization processing.

[0038] S2. Based on the entity feature tag library, obtain the subject sensitivity value and the region sensitivity value; merge the subject sensitivity value and the region sensitivity value to obtain the total entity sensitivity value; obtain the basic public opinion risk value and the public opinion dissemination stage risk value, and after weighted fusion, obtain the total public opinion risk value; match the correction coefficient according to the public opinion dissemination stage, and based on the correction coefficient, merge the total entity sensitivity value and the total public opinion risk value to obtain the comprehensive sensitivity risk value, and after threshold classification, obtain the public opinion risk level.

[0039] First, based on the entity feature label library, the subject sensitivity value and the region sensitivity value are calculated. After weighted fusion, the total entity sensitivity value is obtained, which not only accurately portrays the sensitivity characteristics of the subject itself, but also fully reflects the regional association sensitivity characteristics, effectively improving the objectivity, comprehensiveness and accuracy of the total entity sensitivity value.

[0040] The specific details of the steps for obtaining the subject's sensitivity value are as follows.

[0041] Step 1: Construct an entity relationship graph based on effective public opinion data. In the entity relationship graph, nodes represent core entities and related entities, and edge weights represent the strength of the relationship between entities. The relationship characteristics of core entities and related entities are quantified into visualized and structured graph data, providing accurate correlation basis for subsequent risk diffusion coefficient calculations. This avoids the one-sidedness of calculating sensitivity values ​​based solely on single entity characteristics, making the quantification of risk diffusion characteristics more aligned with the actual dissemination patterns of public opinion.

[0042] Among them, the correlation strength is obtained based on the frequency of co-occurrence between entities in the same text, the frequency of co-occurrence across platforms, and the growth rate of co-occurrence within the same event time window; the growth rate of co-occurrence within the same event time window is the unit time growth rate of the frequency of co-occurrence between two entities within the preset public opinion event time window.

[0043] Step 2: Based on the paths between core entity nodes and associated entity nodes in the entity relationship graph, obtain the risk diffusion coefficient. Specifically, for each core entity node in the entity relationship graph, calculate the shortest path length (preferably the number of connecting edges between nodes) and the edge weight of the shortest path to each associated entity node; normalize the shortest path length by taking its reciprocal, and then weight and fuse it with the edge weight of the shortest path after min-max normalization to obtain the risk diffusion coefficient from the core entity to the associated entities; if the core entity corresponds to multiple associated entities, take the average of all individual risk diffusion coefficients as the final risk diffusion coefficient of the core entity; quantify the diffusion characteristics of the sensitive risks of the core entity through the risk diffusion coefficient, so that the subject sensitivity value not only reflects the sensitivity of the core entity itself, but also takes into account the potential possibility of its risk diffusion.

[0044] Step 3: Multiply the risk diffusion coefficient by the initial subject sensitivity value of the core entity to obtain the subject sensitivity value. By correcting the initial subject sensitivity value with the risk diffusion coefficient, a dual quantification of the core entity's own sensitivity and the possibility of risk diffusion is achieved. This makes the subject sensitivity value more closely reflect the actual sensitivity risks in the public opinion scenario and improves the accuracy of subsequent hierarchical entity sensitivity total value calculations.

[0045] The specific steps for obtaining the initial subject sensitivity value are as follows.

[0046] Step (1): Within the preset public opinion event time window, locate the position of the core entity in the effective public opinion data as the center, and extract text fragments from the preset windows on the left and right to obtain the hit fragments. By extracting text fragments around the core entity, focus on the semantic content directly related to the core entity, avoid interference from irrelevant text in the calculation of sensitive values, and make subsequent sensitive feature matching more accurate. For example: the effective public opinion data text is "XX Province Power Company launched the province's smart electricity management system today, which can realize real-time monitoring of electricity data", the core entity "XX Province Power Company" appears in the 1-8 character position, and the hit fragment of 10 characters on the left and right is "XX Province Power Company launched the province's smart electricity management system today".

[0047] Step (2): Pre-set a sensitive trigger word list for public opinion, such as "not disclosed, trade secret, internal adjustment, not online", etc. Match the hit fragments with the pre-set sensitive trigger word list to obtain the number of times the trigger word is hit, i.e., the trigger proximity. Calculate the ratio of the trigger proximity to the saturation value, and take the minimum value of the ratio and 1. The result is the trigger proximity value. By matching sensitive trigger words, the sensitive semantic features around the core entity are quantified. Combined with the saturation value limit, over-scoring is avoided, so that the quantification of the trigger proximity value is more objective and accurately reflects the degree of association between the core entity and sensitive information.

[0048] Step (3): Based on the position of the core entity in the effective public opinion data, calculate the structural salience value. The earlier the position (e.g., closer to the title or the first paragraph of the body text), the larger the structural salience value. This realizes the quantification of the structural salience of the core entity using the positional features of the public opinion text, making the calculation of the initial sensitivity value more in line with the dissemination pattern of public opinion information. Specifically, preset weights for different positions in the effective public opinion data, such as a weight of 1.0 for the title position, 0.8 for the first paragraph of the body text, 0.5 for the middle paragraph of the body text, and 0.3 for the last paragraph of the body text, locate the position of the core entity in the text, and match the corresponding weight as the structural salience value.

[0049] Step (4): The trigger proximity value and the structural saliency value are fused and corrected by the source credibility coefficient (the source credibility coefficient is obtained based on the reputation of the source platform) to reduce the interference of repeated propagation and low credibility sources and obtain the instance-level initial sensitivity.

[0050] Step (5): To avoid duplicate scoring caused by the reposting of the same article on multiple platforms, calculate the content fingerprint for all public opinion data containing the core entity (using the SimHash algorithm to convert the text into a fixed-length hash value); preset the fingerprint similarity threshold, and classify texts with the same fingerprint or similarity exceeding the threshold into the same duplicate cluster (i.e., texts reposted on multiple platforms of the same article); for each duplicate cluster, select the largest instance-level initial sensitivity value in the cluster as the cluster representative value; truncate and average the representative values ​​of all duplicate clusters, and normalize them by min-max to obtain the initial subject sensitivity value of the core entity, thereby avoiding the duplicate scoring problem caused by the reposting of the same article on multiple platforms and eliminating the interference of redundant data on the initial sensitivity value. For example, the core entity "XX Province Power Company" corresponds to 10 public opinion data texts within a 24-hour time window. After SimHash deduplication and clustering, they are divided into 3 duplicate clusters: Cluster 1 contains 3 reposts from news media, Cluster 2 contains 2 reposts from social media platforms, and Cluster 3 contains 5 reposts from industry forums. The maximum instance-level initial sensitivity values ​​within each cluster are 0.36, 0.25, and 0.30, respectively. After mean aggregation, the initial subject sensitivity value is approximately (0.36 + 0.25 + 0.30) / 3 ≈ 0.303, which is 0.3 after normalization.

[0051] The specific details of the steps for obtaining the region sensitivity value are as follows.

[0052] Step 1: Establish a prior sensitivity database for the region and calculate the prior sensitivity value of the regional entity. The prior sensitivity value is obtained based on the administrative level attribute, key node density attribute, boundary sensitivity attribute, and historical high-risk event attribute. The inherent sensitivity attributes of the region are comprehensively quantified from four dimensions: administrative level, key nodes, boundary characteristics, and historical risks.

[0053] Among them, the administrative level attribute can be preferably achieved by mapping provincial / municipal / district / county levels to different benchmark values; the key node density attribute can be preferably achieved by normalizing and weighting the number of key parks, transportation hubs, and core industrial clusters within the geographical entity according to unit area or population; the boundary sensitivity attribute is obtained by assigning values ​​to border areas, cross-regional channels, and cross-city urban clusters within the geographical entity and then weighting and integrating them; the historical high-risk attribute is obtained based on the proportion of high-risk public opinion, peak frequency, and average peak intensity within the historical period. Step 2: Within the preset time window, collect multi-source attention (sources include platforms, media channels, or collection terminals) for the geographical entity. Attention can be selected from at least one of the following: number of mentions, number of independent accounts, and interaction-weighted mentions; for each source, calculate the historical attention mean and standard deviation, and calculate the logarithmic relative anomaly. Take the average of the logarithmic relative anomalies of all sources as the comprehensive anomaly, and the calculation formula is as follows:

[0054] ,

[0055] For regional entities The abnormality value, For source Standard deviation of attention For source The average level of attention represents the normal level. For source The variance of attention, representing historical fluctuations, To compensate for the value and prevent the denominator from being 0, This represents the total number of sources.

[0056] Step 3: To distinguish between widespread public opinion surges and localized hotspots, calculate the distribution of attention among the child regions of the parent region entity. The parent region entity can be a provincial-level or municipal-level region entity. Based on the attention distribution of the child regions, calculate the hierarchical clustering degree to represent the concentration of public opinion on the region entity. The concentration of lower-level regions is calculated using the following formula:

[0057] ,

[0058] ,

[0059] as sub-region In the parent region The percentage of sub-regions indicates the sub-region The index value in the parent region The proportion of the total indicator value of all sub-regions as sub-region In the source The level of attention, For Father's Region The set of subregions, for example, if If it is a certain province, then It includes all cities under the jurisdiction of the province; if If it is a certain city, then This refers to all districts and counties under the city's jurisdiction. For Father's Region The number of subregions included, i.e., the set Size, For Father's Region Hierarchical clustering , .

[0060] Simultaneously, a regional relationship graph is constructed, where nodes represent regional entities. Edge weights are obtained based on node geographic adjacency (preferably shared boundaries between nodes) and cross-platform propagation coupling (preferably the synchronization rate of nodes for the same event). The comprehensive anomaly degree is diffused based on the edge weights to obtain the spatial spillover degree, which characterizes the regional spillover intensity of risk. The calculation formula is as follows:

[0061] ,

[0062] For regional entities Spatial spillover, For border rights, For region The set of neighboring nodes, This is the Sigmoid function.

[0063] Step 4: Weight and fuse the prior sensitivity value, comprehensive anomaly degree, hierarchical clustering degree, and spatial spillover degree to obtain the regional sensitivity value.

[0064] Then, based on effective public opinion data, the basic risk value of public opinion is obtained and integrated with the risk value of the dissemination stage to obtain the total risk value of public opinion; the risk value of the dissemination stage is obtained based on the dissemination stage of public opinion; the dissemination stage of public opinion includes: germination, fermentation, outbreak, and decline.

[0065] The basic risk value of public opinion is derived from the number of sensitive keywords and accountability trigger words in valid public opinion data. Sensitive keywords can be obtained by querying and matching a preset sensitive keyword table, and accountability trigger words can be obtained by querying and matching a preset accountability trigger word table, such as words like investigation, accountability, violation of discipline, falsification, and leakage.

[0066] The method for obtaining the risk value during the dissemination phase is as follows: Within a preset time window of the public opinion event, a set of indicators used to characterize the dissemination trend of public opinion is statistically analyzed. This set of indicators includes dissemination heat indicators, dissemination growth rate indicators, cross-platform diffusion indicators, and decay indicators. Specifically, the dissemination heat indicator is based on the number of mentions, the number of independent authors, and the number of interactions; the dissemination growth rate indicator is based on the unit-time change rate of dissemination heat between adjacent time windows; the cross-platform diffusion indicator is based on the number of platforms covered or the proportion of simultaneous dissemination across platforms; and the decay indicator is based on the continuous decline in dissemination heat or the intensity of negative growth. The indicator set is then subjected to logarithmic compression and normalization. Subsequently, a weighted fusion calculation was performed to obtain the situation intensity value. The spread heat index, spread growth rate index, and cross-platform diffusion index showed a positive correlation with the situation intensity value, while the decay index showed a negative correlation with the situation intensity value. Based on the index set and preset threshold rules, the stage of public opinion dissemination was determined. Then, a stage mapping coefficient was matched according to the stage of public opinion dissemination, and the situation intensity value was mapped to a coefficient through the stage. After truncation and normalization, the risk value of the dissemination stage was obtained. Among them, the stage mapping coefficient corresponding to the outbreak stage was greater than the stage mapping coefficient corresponding to the fermentation stage, and the stage mapping coefficient corresponding to the fermentation stage was greater than the stage mapping coefficient corresponding to the budding stage or the decline stage.

[0067] Finally, adjustment coefficients are matched according to the stage of public opinion dissemination. Based on these adjustment coefficients, the total sensitivity value of the entity and the total public opinion risk value are merged to obtain a comprehensive sensitivity risk value. This achieves precise adjustment and fusion of the total sensitivity value of the entity and the total public opinion risk value, taking into account both the inherent sensitivity attributes of the entity itself and the dynamic impact of public opinion at different stages of dissemination. Further threshold grading is then applied to obtain the public opinion risk level. The public opinion risk levels include: high risk, medium risk, and low risk. The higher the comprehensive sensitivity risk value, the higher the public opinion risk level.

[0068] This embodiment, based on an entity feature tag library, achieves the fusion and quantification of subject sensitivity values ​​and regional sensitivity values. It obtains a comprehensive sensitivity risk value by weighted fusion of basic public opinion risk values ​​and risk values ​​at the dissemination stage, combined with dynamic adjustment of the dissemination stage correction coefficient. Finally, it completes threshold classification and risk level determination, which can comprehensively assess public opinion risks from multiple dimensions, including the inherent attributes of entities, regional sensitivity characteristics, basic public opinion risks, and dissemination evolution trends, thus significantly improving the accuracy and pertinence of public opinion risk assessment.

[0069] S3. Based on the public opinion risk level and the corresponding stage of public opinion dissemination, the core entities and related entities are anonymized to obtain anonymized public opinion data; the anonymized public opinion data is then input into the public opinion monitoring and analysis model to obtain the anonymized public opinion analysis results.

[0070] First, based on the level of public opinion risk and the corresponding stage of public opinion dissemination, the core entities and related entities in the public opinion data are de-identified to obtain de-identified public opinion data.

[0071] Desensitization processes include irreversible vector mapping replacement, and / or local feature blurring, and / or feature masking.

[0072] The irreversible vector mapping replacement operation involves using a pre-defined irreversible hash vector mapping algorithm to map identity representation information, including entity name, unique identifier, and unique code, into a fixed-dimensional anonymous vector with no semantic association and no reverse restoration path. This ensures that the original entity information cannot be deduced from the mapped vector, achieving complete anonymization and desensitization of the entity's identity. For example, the entity "XX Province Power Company" is replaced with the anonymous vector identifier VEC_7F2D9C through irreversible vector mapping.

[0073] The operation of local feature fuzzification is as follows: Essential pre-defined business characteristics of the entity are retained, including administrative level and category attributes. Pre-defined non-essential sensitive characteristics of the entity, including specific name, detailed address, unique number, and personalized identifier, are generalized and fuzzily replaced, balancing data anonymization effectiveness with business analysis usability. For example, the entity "XX City Electricity Dispatch Center" is fuzzified to "City-level Electricity Dispatch Agency".

[0074] Feature masking involves removing the entity's identifiable and traceable core features, including sensitive attributes, key identity features, and related semantic information (which can be customized as needed). This blocks the propagation and tracing paths of sensitive information, eliminating the risk of sensitive information leakage. For example, the text at the location of the entity "XX Provincial Power Company (responsible for the province's power grid dispatch and power supply guarantee)" is replaced with "[masked]".

[0075] When the public opinion risk level is high and the public opinion dissemination stage is in the nascent or receding stage with a small dissemination scope and controllable impact, the core entities will undergo high-intensity irreversible vector mapping replacement processing to prevent identity leakage, and the related entities will undergo lightweight local feature blurring processing to balance desensitization security and data availability.

[0076] When the public opinion risk level is high and the public opinion dissemination stage is fermentation or outbreak with rapid dissemination, wide coverage, and strong risk spillover, the core entities will be irreversibly replaced by vector mapping and feature masking, and related entities will be irreversibly replaced by vector mapping to completely block the dissemination and tracing of sensitive entity information and minimize the harm of high-risk public opinion information leakage.

[0077] When the public opinion risk level is medium risk and the public opinion dissemination stage is in its infancy or receding, the core entities and related entities will be partially blurred, while retaining the regional administrative level. This lightweight blurring achieves basic desensitization, retains the regional administrative level (regional entity level), and ensures the business functions of regional tracing and hierarchical analysis of public opinion, balancing desensitization compliance with business practicality.

[0078] When the public opinion risk level is medium risk and the public opinion dissemination stage is fermentation or outbreak, the core entity will be replaced by irreversible vector mapping to enhance the desensitization strength of the core entity, and the related entities will be partially blurred. While controlling the desensitization cost, the spread of sensitive information in medium-risk public opinion can be effectively suppressed.

[0079] When the public opinion risk level is low and the public opinion dissemination stage is in its infancy or receding, the original statements and the integrity of the original data should be preserved to maximize the accuracy of public opinion analysis and data retrieval and business efficiency.

[0080] When the public opinion risk level is low and the public opinion dissemination stage is fermentation or outbreak, the core entities and related entities are partially blurred, and only non-core information is blocked. This not only prevents potential information leakage risks, but also fully preserves the data business analysis value, achieving a balance between risk prevention and data utilization.

[0081] Furthermore, to prevent analysis distortion caused by excessive semantic differences between the desensitized data and the original public opinion data, this embodiment, after obtaining the desensitized public opinion data, also includes obtaining the semantic similarity (preferably cosine similarity) between the original public opinion data and the desensitized public opinion data. If the semantic similarity is not less than the semantic similarity threshold, the verification passes; if the semantic similarity is less than the semantic similarity threshold, for entities with partially blurred features, the full name blurring is changed to only individual identifier blurring, while retaining industry, level, and function; for feature-masked fragments, only key identity information is masked, until the semantic similarity of the desensitized public opinion data is not less than the semantic similarity threshold.

[0082] This embodiment implements differentiated de-identification processing for core entities and related entities based on the level of public opinion risk and the stage of dissemination, and generates de-identified public opinion data. The de-identified data is then input into the public opinion monitoring and analysis model for analysis. This can effectively protect sensitive entity information such as key enterprises, local management departments, and key regions throughout the entire process of model entity identification, regional identification, topic judgment, and public opinion early warning. It avoids the leakage or reverse tracing of sensitive identity information in the model input, calculation, and output stages, and achieves full-link data security and privacy protection for the public opinion monitoring and analysis model without affecting the normal analysis capabilities of the model.

[0083] This embodiment also provides a data security protection system, deployed in a public opinion monitoring and analysis model, for implementing the above-mentioned data security protection method, including:

[0084] The entity feature tag library generation module is used to acquire raw public opinion data, remove invalid data, and extract core entities, related entities, and regional entities to form an entity feature tag library; the regional entity levels include provincial regional entities, prefecture-level regional entities, and district / county-level regional entities.

[0085] The public opinion risk level generation module is used to obtain subject sensitivity values ​​and regional sensitivity values ​​based on an entity feature tag library; merge the subject sensitivity values ​​and regional sensitivity values ​​to obtain the total entity sensitivity value; obtain the basic public opinion risk value and the risk value of the public opinion dissemination stage, and obtain the total public opinion risk value through weighted fusion; the public opinion dissemination stages include: budding, fermentation, outbreak, and decline; match the correction coefficient according to the public opinion dissemination stage, and merge the total entity sensitivity value and the total public opinion risk value based on the correction coefficient to obtain the comprehensive sensitivity risk value, and obtain the public opinion risk level through threshold classification;

[0086] The desensitized public opinion analysis result generation module is used to desensitize core entities and related entities according to the public opinion risk level and the corresponding public opinion dissemination stage to obtain desensitized public opinion data. The desensitization process includes irreversible vector mapping replacement, and / or local feature blurring, and / or feature masking. The desensitized public opinion data is input into the public opinion monitoring and analysis model to obtain the desensitized public opinion analysis results.

[0087] This embodiment also provides a data security protection device, including a processor and a memory, wherein the processor implements the above-described data security protection method when executing a computer program stored in the memory.

[0088] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program implements the above-described data security protection method when executed by a processor.

[0089] This embodiment provides a data security protection method applied to the data security protection of a public opinion monitoring and analysis model. First, the original public opinion data is cleaned to remove invalid information, and core entities, related entities, and hierarchical regional entities are extracted to construct an entity feature tag library, laying a standardized and structured foundation for the data security protection of the public opinion monitoring and analysis model. Then, based on this tag library, the total sensitivity value of entities and the total public opinion risk value are quantified. Combined with the public opinion propagation stage, the public opinion risk level is calculated and graded, providing a reliable basis for subsequent precise and differentiated security desensitization. Next, according to the determined public opinion risk level and corresponding propagation stage, irreversible vector mapping replacement, or local feature fuzzification, or feature masking are used for core entities and related entities to achieve differentiated desensitization processing, blocking the leakage of sensitive information and the risk of reverse tracing from the source. Finally, the desensitized public opinion data is input into the public opinion monitoring and analysis model, enabling the model to effectively protect sensitive information during the processing of public opinion data, preventing the leakage or reverse tracing of sensitive identity information, and achieving full-link data security and privacy protection for the public opinion monitoring and analysis model without affecting its normal analytical capabilities.

[0090] This embodiment provides a data security protection method that includes niche related entities and core entities within a unified protection scope. This avoids the risk of leakage of core sensitive entities by reverse deducing from related information, achieving full coverage of public opinion entity protection. Simultaneously, in calculating the comprehensive sensitivity risk value used for public opinion risk level classification, the characteristics of provincial, municipal, and county-level regional entities are considered, and regional sensitivity values ​​are calculated. This solves the imbalance problem of insufficient protection for niche sensitive entities at the county level and excessive desensitization of general regional entities at the provincial level. Furthermore, the protection intensity is dynamically adjusted according to the propagation stages of public opinion—its emergence, fermentation, outbreak, and decline—avoiding the drawbacks of static desensitization affecting handling at the emergence stage and insufficient protection at the outbreak stage, ensuring that security protection is deeply aligned with the development trend of public opinion.

[0091] This embodiment provides a data security protection method that, by constructing a regional prior sensitivity database, quantifying regional sensitivity values, implementing differentiated entity desensitization processing based on public opinion risk levels and dissemination stages, and combining it with multi-dimensional semantic consistency verification, can effectively protect sensitive information such as entities and regions from leakage during model processing, prevent the risk of reverse reconstruction of sensitive information, and ensure the semantic consistency of desensitized data, thereby ensuring the safe and compliant processing of model data and providing comprehensive data security support for the stable and secure operation of public opinion monitoring and analysis models.

[0092] While exemplary embodiments of the invention have been described herein, many other variations or modifications conforming to the principles of the invention can be directly determined or derived from the disclosure of this invention without departing from its spirit and scope. Therefore, the scope of the invention should be understood and recognized to cover all such other variations or modifications.

Claims

1. A data security protection method applied to the data security protection of a public opinion monitoring and analysis model, characterized in that, This includes the following operations: S1. Obtain raw public opinion data, remove invalid data, extract core entities, related entities, and regional entities to form an entity feature tag library; the regional entity levels include provincial regional entities, prefecture-level regional entities, and district / county-level regional entities. S2. Based on the entity feature tag library, obtain the subject sensitivity value and the region sensitivity value. The method for obtaining the subject sensitivity value is as follows: based on effective public opinion data, construct an entity relationship graph with nodes representing core entities and related entities, and edge weights representing the relationship strength between entities; based on the paths between core entity nodes and related entity nodes, obtain the risk diffusion coefficient, and multiply it by the initial subject sensitivity value of the core entity to obtain the subject sensitivity value; the method for obtaining the region sensitivity value is as follows: establish a region prior sensitivity library, calculate the prior sensitivity value of region entities, and the prior sensitivity value is obtained based on administrative level attributes, key node density attributes, boundary sensitivity attributes, and historical high-risk event attributes; within a preset time... Within the framework, multi-source attention is statistically analyzed for regional entities. For each source, the historical mean and standard deviation of attention are calculated, and the log-relative anomaly is determined. The mean of the log-relative anomalies from all sources is taken as the comprehensive anomaly. The attention distribution of child regions for regional entities that are parent regional entities is calculated, and the hierarchical clustering degree is calculated based on the attention distribution of child regional entities. A regional relationship graph is constructed, where nodes are regional entities, and edge weights are obtained based on node geographic adjacency and cross-platform propagation coupling. The comprehensive anomaly is diffused based on the edge weights to obtain the spatial spillover degree. The prior sensitivity value, comprehensive anomaly degree, hierarchical clustering degree, and spatial spillover degree are weighted and fused to obtain the regional sensitivity value. The entity sensitivity value is merged with the regional sensitivity value to obtain the total entity sensitivity value; The basic risk value and the risk value of the dissemination stage of public opinion are obtained, and then weighted and integrated to obtain the total risk value of public opinion. The basic risk value of public opinion is obtained based on the number of sensitive keywords and accountability trigger words in the effective public opinion data; The method for obtaining the risk value during the propagation phase is as follows: within the preset time window of the public opinion event, a set of indicators used to characterize the propagation trend of public opinion is statistically analyzed, logarithmically compressed and normalized, and then weighted and fused to obtain the trend strength value. The situation intensity value is mapped into coefficients according to the stage of public opinion dissemination, and then truncated and normalized to obtain the risk value of the dissemination stage. The stages of public opinion dissemination include: germination, fermentation, outbreak, and decline; Based on the matching correction coefficient of the public opinion dissemination stage, the total sensitivity value of the entity and the total public opinion risk value are integrated to obtain the comprehensive sensitivity risk value. After threshold classification, the public opinion risk level is obtained. S3. Based on the public opinion risk level and the corresponding stage of public opinion dissemination, the core entities and related entities are desensitized to obtain desensitized public opinion data; the desensitization process includes irreversible vector mapping replacement, and / or local feature blurring, and / or feature masking; The anonymized public opinion data is input into the public opinion monitoring and analysis model to obtain the anonymized public opinion analysis results.

2. The data security protection method according to claim 1, characterized in that, In S3, When the public opinion risk level is high and the public opinion dissemination stage is in its infancy or receding, the core entity will be replaced by a high-intensity irreversible vector mapping, and the related entities will be blurred by a lightweight local feature. When the public opinion risk level is high and the public opinion dissemination stage is fermentation or outbreak, the core entities will be replaced by irreversible vector mapping and feature masking, and the related entities will be replaced by irreversible vector mapping. When the public opinion risk level is medium risk and the public opinion dissemination stage is in its infancy or receding, the core entities and related entities will have their local characteristics blurred, while the regional administrative level will be retained. When the public opinion risk level is medium risk and the public opinion dissemination stage is fermentation or outbreak, the core entity will be replaced by irreversible vector mapping and the related entities will be partially blurred. When the public opinion risk level is low and the public opinion dissemination stage is in its infancy or receding, the original statement should be retained. When the public opinion risk level is low and the public opinion dissemination stage is fermentation or outbreak, the core entity and related entities will be partially blurred.

3. The data security protection method according to claim 1, characterized in that, In S3, the irreversible vector mapping replacement operation is as follows: using a preset irreversible hash vector mapping algorithm, the entity's identity representation information is mapped to a fixed-dimensional anonymous vector; the local feature fuzzification operation is as follows: retaining the entity's preset necessary business features, and performing fuzzy replacement processing on the entity's preset local non-essential sensitive features; the feature masking operation is as follows: removing the entity's core features, including sensitive attributes, key identity features, and associated semantic information.

4. The data security protection method according to claim 1, characterized in that, In S3, after obtaining the de-identified public opinion data, it also includes obtaining the semantic similarity between the original public opinion data and the de-identified public opinion data. If the semantic similarity is not less than the semantic similarity threshold, the verification passes; if the semantic similarity is less than the semantic similarity threshold, for entities with blurred local features, the full name blurring is changed to only individual identifier blurring, while retaining industry, level, and function; for feature-masked fragments, only key identity information is masked.

5. The data security protection method according to claim 1, characterized in that, The core entity is the dominant party in the public opinion event within the valid public opinion data; related entities refer to entities that have a direct relationship with the core entity.

6. A data security protection system, deployed in a public opinion monitoring and analysis model, for implementing the data security protection method of claim 1, characterized in that, include: The entity feature tag library generation module is used to acquire raw public opinion data, remove invalid data, and extract core entities, related entities, and regional entities to form an entity feature tag library; the regional entity levels include provincial regional entities, prefecture-level regional entities, and district / county-level regional entities. The public opinion risk level generation module is used to obtain the subject sensitivity value and the region sensitivity value based on the entity feature tag library; The entity sensitivity value is merged with the regional sensitivity value to obtain the total entity sensitivity value; The basic risk value and the risk value of the dissemination stage of public opinion are obtained, and then weighted and integrated to obtain the total risk value of public opinion. The stages of public opinion dissemination include: germination, fermentation, outbreak, and decline; a correction coefficient is matched according to the stages of public opinion dissemination, and based on the correction coefficient, the total sensitivity value of the entity and the total public opinion risk value are integrated to obtain a comprehensive sensitivity risk value, which is then classified by threshold to obtain the public opinion risk level; The desensitized public opinion analysis result generation module is used to desensitize core entities and related entities according to the public opinion risk level and the corresponding public opinion dissemination stage to obtain desensitized public opinion data. The desensitization process includes irreversible vector mapping replacement, and / or local feature blurring, and / or feature masking. The desensitized public opinion data is input into the public opinion monitoring and analysis model to obtain the desensitized public opinion analysis results.

7. A data security protection device, characterized in that, It includes a processor and a memory, wherein the processor implements the data security protection method as described in any one of claims 1-5 when executing a computer program stored in the memory.

8. A computer-readable storage medium, characterized in that, Used for storing computer programs, wherein the computer programs, when executed by a processor, implement the data security protection method as described in any one of claims 1-5.