Data risk management system and method based on large model

Abstract

The invention discloses a data risk management system and method based on a large model, and relates to the technical field of data risk identification, and the data risk management method specifically comprises the following steps: obtaining data; the obtained data is preprocessed; analyzing data by using a large model, constructing a data relation graph, and identifying a potential hidden relation; performing association rule mining to obtain semantic association among the data; obtaining time sequence dependence among the data through dynamic time warping; carrying out correlation analysis through multi-modal fusion by taking the hidden relationship, the semantic association and the time sequence dependence as references; introducing a sensitivity dynamic grading and confidence correction mechanism to fit a risk assessment function; constructing a dynamic threshold function according to time, space and confidence feedback of the large model as elements; and providing coping suggestions according to the risk assessment result. According to the method, the hidden association between the data is identified, semantic analysis and time sequence dependency are combined, a risk assessment function is fitted, and the risk identification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data risk identification, and in particular to a data risk management system and method based on a large model. Background Art

[0002] Data risk identification mainly studies how to identify and evaluate potential risks in data through intelligent analysis, machine learning, deep learning and big data technologies. With the acceleration of digital transformation, enterprises and institutions rely on data-driven decision-making. How to ensure data security and reduce data risks has become an important challenge for information security management.

[0003] With the rapid development of technologies such as big data, artificial intelligence (AI), blockchain and 5G, data-driven decision-making has become increasingly important in the fields of finance, healthcare, industry and the Internet. However, data risks, such as data leakage, data tampering, and abnormal behavior, are becoming increasingly serious. Traditional data risk management methods have certain limitations when facing complex multimodal data. At present, large models (such as GPT, BERT, DALL, etc.) have made significant breakthroughs in natural language processing (NLP), computer vision (CV) and multimodal fusion. Intelligent analysis based on large models can provide more accurate data understanding capabilities and provide more advanced means for data risk management; Most existing technologies use static thresholds, such as fixed credit scores and fixed anomaly detection thresholds, which are difficult to adapt to environmental changes and personalized needs; risk assessment is mostly a comparison of single indicators, without a fusion judgment of multiple indicators. Secondly, the existing confidence calculation method fails to combine historical data and real-time feedback for dynamic adjustment, and has poor adaptability. Summary of the invention

[0004] The purpose of the present invention is to provide a data risk management system and method based on a large model to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solution: a data risk management method based on a large model, the data risk management method specifically comprising the following steps: Acquire data; pre-process the acquired data; Based on the preprocessed data, use a large model to analyze the data, build a data relationship graph through graph neural network technology, and identify potential hidden relationships; Conduct association rule mining to obtain semantic associations between data; Obtain the temporal dependencies between data through dynamic time warping; Based on the hidden relationship, semantic association and temporal dependency, correlation analysis is performed through multimodal fusion; Through the correlation analysis results, a dynamic sensitivity grading and confidence correction mechanism are introduced to fit the risk assessment function; A dynamic threshold function is constructed with time, space and confidence feedback of large models as elements; Provide response suggestions based on risk assessment results; Supports automated decision-making, such as data isolation, permission adjustment, etc.

[0006] Get data, specifically: Provide API interfaces to connect to data, including corporate databases, social media networks, and financial data centers; Data acquisition methods include real-time collection and batch import; Preprocess the acquired data, specifically: The preprocessing includes missing value filling, outlier replacement, data alignment and synchronization, and data standardization; The outlier replacement is performed according to Z-score standardization (3σ principle) to detect outliers and replace them with historical means; according to different actual situations, median filling, mode filling or linear difference etc. can also be selected; Set timestamp for data alignment and synchronization; Specifically: The large model includes text analysis services and multimodal analysis services; The text analysis service is based on the knowledge base comparison results of knowledge base service, adversarial prompt analysis and large model self-consistency verification; The multimodal analysis service is based on the image data storage and metadata analysis to perform image and text consistency results; The method of knowledge base comparison and image-text consistency adopted by the large model is the cosine similarity of space vectors, and the result is mapped to [0,1]; The knowledge base service is a massive text database based on historical data in a large model; The image data storage is a massive image database based on historical data in a large model; The data relationship graph is constructed to identify potential hidden relationships, specifically: The data relationship graph is constructed by using a graph neural network based on the knowledge base comparison result and the image-text consistency result; The data relationship diagram is used to model the association relationship between data; The data relationship graph includes nodes and edges; The nodes represent data entities, including users, devices, data records, texts, and images; The edges represent the relationships between the data entities, including data sources, transaction paths, social associations, and access records; According to the constructed data relationship diagram, the association relationship between data entities is obtained.

[0007] Node attributes represent the characteristics of an entity, such as user credit score, transaction amount, text sentiment score, etc.

[0008] Edge attributes represent the characteristics of the relationship, such as the number of transactions, data similarity, access rights, etc. Based on the hidden relationship, semantic association and temporal dependency, correlation analysis is performed through multimodal fusion, specifically: Step S6-1, based on the big model, real-time calculation of the semantic association between data entities; Step S6-2, obtaining the shortest path weight between data entities through the constructed data relationship graph; Step S6-3, obtaining the time series correlation between data entities through dynamic time warping; Step S6-4, multi-modal fusion, fitting correlation analysis function, and calculating multi-dimensional dynamic correlation; The specific representation formula of the multi-dimensional dynamic correlation is: ; in, Represents the multi-dimensional dynamic correlation between data entities A and B; Indicates the semantic association between data entities A and B, obtained based on the large model; Represents the shortest path weight between data entities A and B, obtained based on the data relationship graph; Indicates the time series correlation between data entities A and B; , and Represents the weight coefficient, satisfying the constraint + + =1; The dynamic sensitivity classification and confidence correction mechanism are introduced to fit the risk assessment function, specifically: According to the multi-dimensional dynamic correlation between real-time data, taking the big model as the benchmark, the authenticity score and confidence of the current data entity in the big model are obtained, and the risk assessment value of the current data entity is calculated in real time; Specifically: ; in, Indicates the risk assessment value of the current data entity X; Indicates the maximum value of the multi-dimensional dynamic correlation between the current data entity X and other data entities; Indicates the authenticity score of the current data entity X, obtained by the large model; Represents the confidence of the current data entity X, obtained by the large model; Indicates the sensitivity of the current data entity X, obtained from the preset data comparison table; The authenticity score is calculated based on cosine similarity; A dynamic threshold function is constructed for the features based on the confidence feedback of time, space and large models, specifically: The dynamic threshold function includes a basic threshold, a contextual environment perception influence, and a reinforcement learning dynamic correction value; Obtain the risk assessment value of the data entity in history, and calculate the mean and standard deviation of the risk assessment value; calculate the basic threshold based on the number of high-risk alarms in history and the historical average traffic level; The impact of contextual environment perception is calculated based on the mean confidence value of the big model in judging the authenticity of the data and the edge density of the data relationship graph constructed in real time; The dynamic correction value of reinforcement learning is calculated based on the historical false alarm rate and the recall rate target preset by the large model; The formula representation of the dynamic threshold function is specifically: ; in, represents the dynamic threshold at time t; t represents the time variable; represents the basic threshold; represents the environmental complexity factor at time t; represents the false positive penalty factor; ; in, Represents the average risk assessment value of the data entity in history; Represents the standard deviation of the risk assessment value of the data entity in history; Indicates the number of high-risk alarms in the same period of history; Indicates the ratio of current data traffic to the historical average traffic level; ; in, Indicates the mean confidence level of the big model in judging the authenticity of the current data segment; Represents the edge density of the data relationship graph; Calculate the edge density of graph neural networks (number of edges / number of nodes²); ; in, It represents the false alarm rate in the past period of time at time t; Indicates the preset minimum recall rate target of the large model; A data risk management system based on a big model, the data risk management system comprising a data acquisition and preprocessing module, a big model deployment module, a data analysis module, a risk assessment module, a dynamic threshold module and a visual interaction module; The data acquisition and preprocessing module is used to obtain data through the API interface, supporting real-time acquisition and batch import; The preprocessing includes missing value filling, outlier detection and replacement, and data standardization; The large model deployment module is used to provide knowledge reasoning, text understanding and multimodal data fusion capabilities based on big data, and provide intelligent support for data analysis; including text analysis services and multimodal analysis services; The text analysis service includes knowledge base comparison, adversarial prompt analysis and self-consistency verification; The multimodal analysis service includes image data storage and analysis and image-text consistency analysis; The data analysis module is used to mine association rules based on the automated comparative analysis results of the large model, obtain semantic associations between data, obtain temporal dependencies between data through dynamic time warping, and perform correlation analysis through multimodal fusion; The risk assessment module is used to calculate the risk assessment value of the current data entity in real time based on the multi-dimensional dynamic correlation, the authenticity score and the confidence of the large model; The dynamic threshold module is used to dynamically adjust the threshold through reinforcement learning by referring to the mean and standard deviation of historical risk assessment values ​​and combining the environmental complexity factor and the false alarm penalty factor; The visualization interaction module is used to provide visualization of risk assessment results and display dynamic threshold change curves; Provide feedback on data analysis results; Support automated decision-making operations (data isolation, permission adjustment); Configure warning strategies to issue warnings for data that exceeds dynamic thresholds calculated in real time.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention is based on large models, graph neural networks (GNNs) and multimodal analysis, combined with a dynamic threshold mechanism, to achieve more accurate, efficient and intelligent data risk management; by constructing a data relationship graph, identifying hidden associations between data, combining semantic analysis and time series dependencies, fitting risk assessment functions, and improving the accuracy of risk identification; adopting dynamic sensitivity grading and confidence correction mechanisms to make risk assessment more flexible and adaptive; adjusting dynamic thresholds based on reinforcement learning to reduce false alarm rates and missed alarm rates and optimize the alarm system; at the same time, supporting automated decision-making to achieve intelligent responses such as data isolation and authority adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic diagram of a flow chart of a data risk management system and method based on a large model of the present invention; Figure 2 A schematic diagram of a big model structure of a data risk management system and method based on a big model of the present invention; Figure 3 This is a schematic diagram of data relationships in Example 2 of a data risk management system and method based on a large model of the present invention. DETAILED DESCRIPTION

[0011] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0012] Example: Figure 1-Figure 3 As shown, the present invention provides a technical solution, a data risk management method based on a large model, such as Figure 1 As shown, the data risk management method specifically includes the following steps: Acquire data; pre-process the acquired data; Based on the preprocessed data, use a large model to analyze the data, build a data relationship graph through graph neural network technology, and identify potential hidden relationships; Conduct association rule mining to obtain semantic associations between data; Obtain the temporal dependencies between data through dynamic time warping; Based on the hidden relationship, semantic association and temporal dependency, correlation analysis is performed through multimodal fusion; Through the correlation analysis results, a dynamic sensitivity grading and confidence correction mechanism are introduced to fit the risk assessment function; A dynamic threshold function is constructed with time, space and confidence feedback of large models as elements; Provide response suggestions based on risk assessment results; Supports automated decision-making, such as data isolation, permission adjustment, etc.

[0013] Get data, specifically: Provide API interfaces to connect to data, including corporate databases, social media networks, and financial data centers; Data acquisition methods include real-time collection and batch import; Preprocess the acquired data, specifically: The preprocessing includes missing value filling, outlier replacement, data alignment and synchronization, and data standardization; The outlier replacement is performed according to Z-score standardization (3σ principle) to detect outliers and replace them with historical means; according to different actual situations, median filling, mode filling or linear difference etc. can also be selected; Set timestamps for data alignment and synchronization; Specifically: like Figure 2 As shown, the large model includes text analysis services and multimodal analysis services; The text analysis service is based on the knowledge base comparison results of knowledge base service, adversarial prompt analysis and large model self-consistency verification; The multimodal analysis service is based on the image data storage and metadata analysis to perform image and text consistency results; The method of knowledge base comparison and image-text consistency adopted by the large model is the cosine similarity of space vectors, and the result is mapped to [0,1]; The knowledge base service is a massive text database based on historical data in a large model; The image data storage is a massive image database based on historical data in a large model; The data relationship graph is constructed to identify potential hidden relationships, specifically: The data relationship graph is constructed by using a graph neural network based on the knowledge base comparison result and the image-text consistency result; The data relationship diagram is used to model the association relationship between data; The data relationship graph includes nodes and edges; The nodes represent data entities, including users, devices, data records, texts, and images; The edges represent the relationships between the data entities, including data sources, transaction paths, social associations, and access records; According to the constructed data relationship diagram, the association relationship between data entities is obtained.

[0014] Node attributes represent the characteristics of an entity, such as user credit score, transaction amount, text sentiment score, etc.

[0015] Edge attributes represent the characteristics of the relationship, such as the number of transactions, data similarity, access rights, etc. Based on the hidden relationship, semantic association and temporal dependency, correlation analysis is performed through multimodal fusion, specifically: Step S6-1, based on the big model, real-time calculation of the semantic association between data entities; Step S6-2, obtaining the shortest path weight between data entities through the constructed data relationship graph; Step S6-3, obtaining the time series correlation between data entities through dynamic time warping; Step S6-4, multi-modal fusion, fitting correlation analysis function, and calculating multi-dimensional dynamic correlation; The specific representation formula of the multi-dimensional dynamic correlation is: ; in, Represents the multi-dimensional dynamic correlation between data entities A and B; Indicates the semantic association between data entities A and B, obtained based on the large model; Represents the shortest path weight between data entities A and B, obtained based on the data relationship graph; Indicates the time series correlation between data entities A and B; , and Represents the weight coefficient, satisfying the constraint + + =1; The dynamic sensitivity classification and confidence correction mechanism are introduced to fit the risk assessment function, specifically: According to the multi-dimensional dynamic correlation between real-time data, taking the big model as the benchmark, the authenticity score and confidence of the current data entity in the big model are obtained, and the risk assessment value of the current data entity is calculated in real time; Specifically: ; in, Indicates the risk assessment value of the current data entity X; Indicates the maximum value of the multi-dimensional dynamic correlation between the current data entity X and other data entities; Indicates the authenticity score of the current data entity X, obtained by the large model; Represents the confidence of the current data entity X, obtained by the large model; Indicates the sensitivity of the current data entity X, obtained from the preset data comparison table; A dynamic threshold function is constructed for the features based on the confidence feedback of time, space and large models, specifically: The dynamic threshold function includes a basic threshold, a contextual environment perception influence, and a reinforcement learning dynamic correction value; Obtain the risk assessment value of the data entity in history, and calculate the mean and standard deviation of the risk assessment value; calculate the basic threshold based on the number of high-risk alarms in history and the historical average traffic level; The impact of contextual environment perception is calculated based on the mean confidence value of the big model in judging the authenticity of the data and the edge density of the data relationship graph constructed in real time; The dynamic correction value of reinforcement learning is calculated based on the historical false alarm rate and the recall rate target preset by the large model; The formula representation of the dynamic threshold function is specifically: ; in, represents the dynamic threshold at time t; t represents the time variable; represents the basic threshold; represents the environmental complexity factor at time t; represents the false positive penalty factor; ; in, Represents the average risk assessment value of the data entity in history; Represents the standard deviation of the risk assessment value of the data entity in history; Indicates the number of high-risk alarms in the same period of history; Indicates the ratio of current data traffic to the historical average traffic level; ; in, Indicates the mean confidence level of the big model in judging the authenticity of the current data segment; Represents the edge density of the data relationship graph; Calculate the edge density of graph neural networks (number of edges / number of nodes²); ; in, It represents the false alarm rate in the past period of time at time t; Indicates the preset minimum recall rate target of the large model; False positive results marked by manual review are entered into the training pool in real time; Update the penalty factor every 30 minutes; Example 1, transaction data analysis: connect to a company A database through an API interface; obtain data: transaction ID: TX20231115103******; initiating account: U****** (personal account); receiving account: CORP_***** (corporate account); amount: $******; timestamp: 202*-11-15 10:32:47.856 GMT+8; transaction IP: 192.168.***.*** (physical location: Malay**ji**); device fingerprint: MAC: 00-1A-3F-**-**-** (number of historical associated accounts: 3); Data preprocessing: time-space alignment, found that the distance deviation between IP location and account registration location is 3,200 kilometers; calculated reasonable moving speed: time difference = 2 hours (last login in * sea); theoretical achievable speed = 3,200 / 2 = 1,600km / h (supersonic passenger plane speed); Calculate text relevance using cosine similarity (large model): Extract the description of the business account: "Import and export trading company, mainly engaged in electronic components" Knowledge base comparison (HS code library): Keywords-knowledge base matching; Electronic components - 72%; import and export trade - 88%; ; Scanned copy of business license vs customs filing information: Image feature similarity: 0.65; Text consistency score: 0.32; ; Construct a transaction subgraph: Node properties: Node ID type credit score Account Age U****** personal 680 3 years CORP_******** enterprise 420 6 months Table 1 Edge properties: Relationship-number of transactions-total amount; Transfer-First-time-$******; Shortest path weight calculation: It is found that CORP_***** is associated with the marked fraud account FRAUD_001 through 4 layers: According to the time relationship, calculation is performed: ; Time Series Analysis: Compare the CORP_***** account fund inflow / outflow pattern (DTW algorithm): ; Adjust factors based on real-time threat intelligence: α=0.5 (semantic weight), β=0.3 (graph relation weight), γ=0.2 (temporal weight); ; Sensitivity rating: According to the relationship table between account type and amount sensitivity: Account type - amount threshold - sensitivity Sen (X); Individual VIP-$100,000-**; Risk assessment value calculation: ; Dynamic Threshold Adjustment: Basic threshold calculation: ; Environmental factor calculation: ; ; Decision trigger: Current risk assessment value 1.189 < dynamic threshold 2.88; marked as medium risk; Automated measures: Trigger facial recognition verification.

[0016] A data risk management system based on a big model, the data risk management system comprising a data acquisition and preprocessing module, a big model deployment module, a data analysis module, a risk assessment module, a dynamic threshold module and a visual interaction module; The data acquisition and preprocessing module is used to obtain data through the API interface, supporting real-time acquisition and batch import; The preprocessing includes missing value filling, outlier detection and replacement, and data standardization; The large model deployment module is used to provide knowledge reasoning, text understanding and multimodal data fusion capabilities based on big data, and provide intelligent support for data analysis; including text analysis services and multimodal analysis services; The text analysis service includes knowledge base comparison, adversarial prompt analysis and self-consistency verification; The multimodal analysis service includes image data storage and analysis and image-text consistency analysis; The data analysis module is used to mine association rules based on the automated comparative analysis results of the large model, obtain semantic associations between data, obtain temporal dependencies between data through dynamic time warping, and perform correlation analysis through multimodal fusion; The risk assessment module is used to calculate the risk assessment value of the current data entity in real time based on the multi-dimensional dynamic correlation, the authenticity score and the confidence of the large model; The dynamic threshold module is used to dynamically adjust the threshold through reinforcement learning by referring to the mean and standard deviation of historical risk assessment values ​​and combining the environmental complexity factor and the false alarm penalty factor; The visualization interaction module is used to provide visualization of risk assessment results and display dynamic threshold change curves; Provide feedback on data analysis results; Support automated decision-making operations (data isolation, permission adjustment); Configure warning strategies to issue warnings for data that exceeds dynamic thresholds calculated in real time.

[0017] Embodiment 2, social media data management; During the public health incident at ** University (a high incidence of influenza), a social platform detected 1.2 million new related posts in a single day, which were suspected to include: False preventive measures (e.g., "drinking strong alcohol can kill the virus"); Drug fraud information (selling certain special medicines); Multi-source data access: Data Types Collection frequency Data volume example Post content real time 5,00 pieces / second (including pictures / videos) User Conduct real time Like / Repost / Comment data 2,000 times / second Device fingerprint batch 200,000 device ID associations updated daily Table 2 External data (updated): Official knowledge base of the National Health Commission (updated every 30 minutes); Cyberspace Administration rumor sample database (including 12,000 historical rumor features); Abnormal account detection: Case sample: Account @HealthGuard**** posted 82 "special medicine" ads within 3 hours after registration; Z-score calculation: According to historical data, the average number of posts posted by a normal user is 3 per hour; σ = 2.1; This user's z=(82-3) / 2.1=37.6; serious abnormality; Handling measures: Temporarily limit the posting frequency to 5 posts per hour; trigger the manual review queue; Among them, the key sentences are: "A certain organization will spray air purifiers starting tomorrow"; Knowledge base comparison: Dimensions - official information - text to be checked - similarity; Time - No plan - "tomorrow" - 0%; Measures - conventional disinfection - "city-wide spraying" - 12%; ; like Figure 3 As shown, node relationship discovery and device fingerprint association: 6 accounts share the same device ID: IMEI358901045******; The MAC address randomization pattern is completely consistent (changed 3 times per hour); Behavioral pattern analysis:

[0018] Table 3 ; Obtained based on device association and copywriting similarity; Multidimensional correlation analysis, time series analysis: It was detected that the water army group concentrated on posting between 22:00 and 24:00 (normal user activity decreased by 30%); DTW distance calculation shows that the similarity with the historical rumor propagation pattern is 0.91; Dynamic weight allocation (during hot events): α=0.6 (semantic weight), β=0.25 (network relationship), γ=0.15 (time series); ; RDS=5.19; Environmental parameter perception: Hot Event Index: ; The basic threshold is 4.0; the training model false alarm rate is 8%; the target recall rate is 90%; Calculate the real-time threshold: DRT=8.01; Risk level determination: Current RDS=5.19<dynamic threshold DRT=8.01; marked as medium-high risk; Disposal strategy: limit flow; add "content in doubt" label; push official rumor-refuting information to interactive users.

[0019] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A data risk management method based on a large model, characterized by: The data risk management method specifically includes the following steps: Acquire data; pre-process the acquired data; Based on the preprocessed data, use a large model to analyze the data, build a data relationship diagram, and identify potential hidden relationships; Conduct association rule mining to obtain semantic associations between data; Obtain the temporal dependencies between data through dynamic time warping; Based on the hidden relationship, semantic association and temporal dependency, correlation analysis is performed through multimodal fusion; Introduce dynamic sensitivity grading and confidence correction mechanism to fit the risk assessment function; A dynamic threshold function is constructed with time, space and confidence feedback of large models as elements; Provide response suggestions based on the risk assessment results.

2. The data risk management method based on a large model according to claim 1, characterized in that: Get data, specifically: Provide API interfaces to connect to data, including corporate databases, social media networks, and financial data centers; Data acquisition methods include real-time collection and batch import.

3. The data risk management method based on a large model according to claim 2 is characterized by: Preprocess the acquired data, specifically: The preprocessing includes missing value filling, outlier replacement, data alignment and synchronization, and data standardization; The outlier replacement is performed based on Z-score standardization to detect outliers and replace them using historical means; Set timestamp for data alignment and synchronization.

4. The data risk management method based on a large model according to claim 3 is characterized by: Specifically: The large model includes text analysis services and multimodal analysis services; The text analysis service is based on the knowledge base comparison results of knowledge base service, adversarial prompt analysis and large model self-consistency verification; The multimodal analysis service is based on the image data storage and metadata analysis to perform image and text consistency results.

5. The data risk management method based on a large model according to claim 4 is characterized in that: The data relationship graph is constructed to identify potential hidden relationships, specifically: The data relationship graph is constructed by using a graph neural network based on the knowledge base comparison result and the image-text consistency result; The data relationship diagram is used to model the association relationship between data; The data relationship graph includes nodes and edges; The nodes represent data entities, including users, devices, data records, texts, and images; The edges represent the relationships between the data entities, including data sources, transaction paths, social associations, and access records; According to the constructed data relationship diagram, the association relationship between data entities is obtained.

6. The data risk management method based on a large model according to claim 5 is characterized by: Based on the hidden relationship, semantic association and temporal dependency, correlation analysis is performed through multimodal fusion, specifically: Step S6-1, based on the big model, real-time calculation of the semantic association between data entities; Step S6-2, obtaining the shortest path weight between data entities through the constructed data relationship graph; Step S6-3, obtaining the time series correlation between data entities through dynamic time warping; Step S6-4: multimodal fusion, fitting correlation analysis function, and calculating multidimensional dynamic correlation.

7. The data risk management method based on a large model according to claim 6 is characterized by: The dynamic sensitivity classification and confidence correction mechanism are introduced to fit the risk assessment function, specifically: According to the multi-dimensional dynamic correlation between real-time data, taking the big model as the benchmark, the authenticity score and confidence level of the current data entity in the big model are obtained, and the risk assessment value of the current data entity is calculated in real time.

8. The data risk management method based on a large model according to claim 7 is characterized by: A dynamic threshold function is constructed for the features based on the confidence feedback of time, space and large models, specifically: The dynamic threshold function includes a basic threshold, a contextual environment perception influence, and a reinforcement learning dynamic correction value; Obtain the risk assessment value of the data entity in history, and calculate the mean and standard deviation of the risk assessment value; calculate the basic threshold based on the number of high-risk alarms in history and the historical average traffic level; The impact of contextual environment perception is calculated based on the mean confidence value of the big model in judging the authenticity of the data and the edge density of the data relationship graph constructed in real time; The dynamic correction value of reinforcement learning is calculated based on the historical false alarm rate and the recall rate target preset by the large model.

9. A data risk management system based on a large model, applying a data risk management method based on a large model as claimed in any one of claims 1 to 8, characterized in that: The data risk management system includes a data acquisition and preprocessing module, a large model deployment module, a data analysis module, a risk assessment module, a dynamic threshold module and a visual interaction module; The data acquisition and preprocessing module is used to obtain data through the API interface, supporting real-time acquisition and batch import; The preprocessing includes missing value filling, outlier detection and replacement, and data standardization; The large model deployment module is used to provide knowledge reasoning, text understanding and multimodal data fusion capabilities based on big data, and provide intelligent support for data analysis; including text analysis services and multimodal analysis services; The text analysis service includes knowledge base comparison, adversarial prompt analysis and self-consistency verification; The multimodal analysis service includes image data storage and analysis and image-text consistency analysis; The data analysis module is used to mine association rules based on the automated comparative analysis results of the large model, obtain semantic associations between data, obtain temporal dependencies between data through dynamic time warping, and perform correlation analysis through multimodal fusion; The risk assessment module is used to calculate the risk assessment value of the current data entity in real time based on the multi-dimensional dynamic correlation, the authenticity score and the confidence of the large model; The dynamic threshold module is used to dynamically adjust the threshold through reinforcement learning by referring to the mean and standard deviation of historical risk assessment values ​​and combining the environmental complexity factor and the false alarm penalty factor; The visualization interaction module is used to provide visualization of risk assessment results and display dynamic threshold change curves; Provide feedback on data analysis results; Configure warning strategies to issue warnings for data that exceeds dynamic thresholds calculated in real time.

Images