A method and system for intelligent security risk assessment based on multi-source data fusion
By standardizing and calibrating the credibility of multi-source security data, constructing a multi-source security relationship diagram, and utilizing an improved GATNE model, the consistency and false alarm issues in multi-source security data fusion were resolved, achieving accuracy and traceability in risk assessment.
Patent Information
- Application Number
- CN202611236798.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies suffer from issues such as data inconsistency, timestamp deviation, duplicate records, false alarm records, and alarm conflicts in multi-source security data fusion, making it difficult to accurately assess risks and trace the source of events, and lacking the ability to express the linkage and impact between objects, areas, and events.
By collecting multi-source security data, performing standardized processing and credibility calibration, a multi-source security relationship diagram is constructed. The improved GATNE model is used to generate risk representation vectors, which are written back to object, event, and area nodes to generate a risk transmission relationship table, tracing the initial event node and related area nodes.
It improves the accuracy and interpretability of risk assessment, reduces false alarm interference, accurately locates risk objects and areas, and provides reliable judgment on disposal priorities.
Smart Images

Figure CN123070701A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Internet of Things (IoT) security technology, and in particular to an intelligent security risk assessment method and system based on multi-source data fusion. Background Technology
[0002] With the development of smart building, park security, and urban comprehensive governance systems, various types of security data, such as video surveillance, access control, perimeter alarms, fire detection, environmental sensing, and security equipment status monitoring, are widely integrated into unified platforms for comprehensive assessment of personnel, vehicle, area, and event risks. Existing technologies typically collect multi-source security data and perform format standardization, time alignment, and alarm aggregation to identify abnormal events and determine risk levels. Some methods further incorporate machine learning models or graph structure analysis methods to fuse target records, area status, and alarm information to improve the automation level of security risk assessment.
[0003] In existing technologies, multi-source security data typically originates from different devices, systems, and sampling periods. This leads to issues such as inconsistent field standards, timestamp discrepancies, duplicate records, missed records, and false alarms. Current risk assessment methods often directly aggregate or merge various alarm results with fixed weights, making it difficult to calibrate the reliability of different data sources using historical false alarm data. Furthermore, it's challenging to utilize consistency between records from different sources within the same object, area, and time period to correct abnormal alarms. This results in false alarms, missed alarms, and alarm conflicts significantly impacting the final assessment results.
[0004] Meanwhile, existing technologies, when constructing security risk correlations, often focus on single device alarms or single-area event statistics, lacking a unified expression of heterogeneous relationships between object nodes, area nodes, and event nodes. They also underutilize trusted relationship edges and area constraint edges, making it difficult to reflect the interconnected impacts of object movement, area sensitivity levels, access permissions, deployment time periods, and device coverage. Even when using graph neural networks or embedded models for risk characterization, the model output is often directly used for classification or scoring, lacking a process of writing risk characterization back to nodes and forming risk transmission relationships along relationship edges that include event time periods. This makes it difficult to accurately trace risk propagation paths, initial event nodes, and associated area nodes.
[0005] Therefore, how to provide an intelligent security risk assessment method and system based on multi-source data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose an intelligent security risk assessment method and system based on multi-source data fusion. This invention fully utilizes multi-source security data fusion, data source credibility calibration, and an improved GATNE model. It details the process of intelligently realizing standardized processing of security data, fusion of credible security events, construction of multi-source security relationship diagrams, extraction of risk representations, risk transmission and location, and risk level assessment. It has the advantages of low false alarm interference, accurate risk location, strong event tracing capability, and reliable judgment of handling priorities.
[0007] An intelligent security risk assessment method based on multi-source data fusion according to an embodiment of the present invention includes:
[0008] Collect multi-source security data, historical false alarm data, and security constraint data, and preprocess them to generate standard security data;
[0009] Based on historical false alarm data and cross-source consistency calibration standard security data, generate trusted security data and data source credibility;
[0010] Merge target records, area records, and alarm records from trusted security data to generate trusted security event data;
[0011] A multi-source security relationship graph is constructed based on trusted security event data, data source credibility, and security constraint data. The multi-source security relationship graph includes object nodes, region nodes, event nodes, trusted relationship edges, and region constraint edges.
[0012] The multi-source security relationship graph is input into the improved GATNE model to generate a risk representation vector. The improved GATNE model includes a security object attribute embedding layer, a trusted relationship edge embedding calibration layer, a relationship type attention fusion layer, a multi-granularity security risk projection layer, and a heterogeneous Skip-Gram output layer.
[0013] Write the risk representation vector back to the object node, event node, and region node, and generate a risk transmission relationship table containing the event time period along the trusted relationship edge and the region constraint edge;
[0014] Based on the risk transmission relationship table and the credibility of the data source, the conflict of multiple alarms from the same object, the same region, and the same time period is corrected, the first event node and related regional nodes are traced, and the risk location results are generated.
[0015] Based on the risk location results and security constraint data, a security risk assessment result is generated, which includes the risk object, risk area, risk type, risk level, and handling priority.
[0016] Optionally, the preprocessing to generate standard security data includes:
[0017] The data items for multi-source security data, historical false alarm data, and security constraint data are identified, and source identifier completion, field name unification, data format conversion, and anomaly sampling and removal are performed. The multi-source security data includes video recognition data, access control data, perimeter alarm data, fire sensing data, environmental sensing data, and security equipment status data. The historical false alarm data includes false alarm event identifier, false alarm source device, false alarm occurrence area, false alarm occurrence time, false alarm trigger type, and false alarm handling result. The security constraint data includes area sensitivity level, access permission constraint, deployment time period constraint, equipment coverage range constraint, and risk handling level constraint.
[0018] Based on the collection time, device timestamp, alarm trigger time, and handling time, the time benchmark is unified and the event time period is segmented to generate time-series security data;
[0019] Based on object identifier, area number, device number, event type, and constraint type, associated time-series security data, historical false alarm data, and security constraint data, perform missing field filling, duplicate record merging, and sensitive field desensitization to generate standard security data.
[0020] Optionally, the credibility of the generated trusted security data and the data source includes:
[0021] Based on historical false alarm data, the number of false alarms, the number of valid alarms, and the number of times the handling results are consistent are counted according to the source device, event type, occurrence area, and occurrence time, and a historical false alarm feature table is generated.
[0022] Perform event matching on video recognition records, access control records, perimeter alarm records, and environmental perception records for the same target, the same area, and the same time period in standard security data to generate cross-source consistent records;
[0023] The data source credibility is obtained by weighted summation of historical false alarm suppression score, cross-source consistency score, complete collection score and time synchronization score. Based on the data source credibility, the standard security data is then subjected to credibility marking, conflict record downweighting and low credibility record removal to generate credible security data.
[0024] Optionally, the generation of trusted security event data includes:
[0025] Candidate security event records are generated by combining target records, area records, and alarm records according to target identifier, area number, device number, alarm type, and event time period.
[0026] The event merge score is obtained by subtracting the time interval penalty value from the weighted result of target similarity, region overlap, alarm type consistency and data source credibility.
[0027] Candidate security event records that meet the merging threshold are grouped into the same trusted security event. The target record, area record, alarm record, source device, and event time period are merged to generate trusted security event data.
[0028] Optionally, the construction of the multi-source security relationship diagram based on trusted security event data, data source credibility, and security constraint data includes:
[0029] Map target records, area records, and alarm records to object nodes, area nodes, and event nodes, configure node types, node sources, and event time periods, and generate a security node table;
[0030] Match the relationships between object nodes, region nodes, and event nodes to generate trusted relationship edges;
[0031] Generate area constraint edges based on area sensitivity level, access permission constraints, deployment time constraints, and equipment coverage constraints;
[0032] Encapsulate object nodes, region nodes, event nodes, trusted relationship edges, and region constraint edges to generate a multi-source security relationship graph.
[0033] Optionally, the generated risk representation vector includes:
[0034] Based on the GATNE model's basic embedding, edge embedding, global node embedding generation, and heterogeneous Skip-Gram output, an improved GATNE model is formed by adding a security object attribute embedding layer at the basic embedding input, a trusted relationship edge embedding calibration layer at the edge embedding branch, a relationship type attention fusion layer at the edge type self-attention combination position, and a multi-granularity security risk projection layer at the global node embedding generation output. The training objects of the heterogeneous Skip-Gram output layer are set as object nodes, region nodes, and event nodes.
[0035] Input the multi-source security relationship diagram into the security object attribute embedding layer, and generate a security node embedding matrix according to node type, node source, event time period and node attribute;
[0036] The security node embedding matrix, trusted relation edges, regional constraint edges, and edge credibility are input into the trusted relation edge embedding calibration layer, and a trusted edge embedding matrix is generated according to edge type, edge connection node, edge credibility, and constraint label.
[0037] Input the security node embedding matrix and the trusted edge embedding matrix into the relation type attention fusion layer, and generate the relation fusion node embedding matrix according to the relation type of trusted relation edges and region constraint edges;
[0038] The relationship fusion node embedding matrix is input into the multi-granularity security risk projection layer, and a multi-granularity risk embedding matrix is generated according to object granularity, region granularity and event granularity.
[0039] The multi-granularity risk embedding matrix is input into the heterogeneous Skip-Gram output layer, and the context node prediction results are generated based on the node walk sequence. The multi-granularity risk embedding matrix is then aggregated to generate a risk representation vector.
[0040] The improved GATNE model is trained by weighting the context node prediction error, the credible relation edge reconstruction error, and the multi-granularity risk embedding consistency error, so that the trained improved GATNE model can be used to generate risk representation vectors.
[0041] Optionally, generating the risk transmission relationship table containing event time periods along the trusted relationship edges and regional constraint edges includes:
[0042] The risk representation vector is written back to the object node, event node, and region node according to the node identifier, generating a node risk write-back table;
[0043] Traverse trusted relation edges and region constraint edges, combine edge connection direction, edge type, edge trustworthiness and constraint label to generate candidate risk transmission records;
[0044] Calculate the risk transmission score of the candidate risk transmission record. The risk transmission score is obtained by subtracting the event time interval penalty value from the weighted average of the risk representation similarity, edge credibility and constraint matching value between the starting node and the ending node.
[0045] Candidate risk transmission records that meet the transmission threshold are selected and organized into a risk transmission relationship table containing the event time period.
[0046] Optionally, the generated risk location results include:
[0047] Multi-source alarm records in the risk transmission relationship table are aggregated by object node, region node, and event time period, and alarm conflict records within the same object node, the same region node, and the same event time period are marked.
[0048] By combining the data source credibility and risk transmission relationship table, a reliable correction is performed on the alarm conflict records to generate corrected alarm records;
[0049] Backtrack along trusted relation edges and region constraint edges to correct the event nodes corresponding to alarm records and determine the first event node;
[0050] Associate the initial event node, object node, and region node to generate a risk location result that includes the risk object, the initial event node, the associated region node, and the location time period.
[0051] Optionally, generating a security risk assessment result that includes the risk object, risk area, risk type, risk level, and handling priority includes:
[0052] Match the risk object, the initial event node, the associated regional nodes and the location time period with the regional sensitivity level, access permission constraints, deployment time constraints and risk handling level constraints to generate a basic record for risk assessment;
[0053] The event type corresponding to the first event node is determined as the risk type, and the area location corresponding to the associated area node is determined as the risk area, and a risk assessment association record is generated.
[0054] Risk levels are determined by risk type, regional sensitivity level, access control constraints, and deployment time constraints; and disposal priorities are determined by risk level and risk disposal level constraints.
[0055] Combine the risk object, risk area, risk type, risk level, handling priority, and location time period to generate a security risk assessment result.
[0056] An intelligent security risk assessment system based on multi-source data fusion according to an embodiment of the present invention includes the following modules:
[0057] The data standardization module is used to process multi-source security data, historical false alarm data, and security constraint data to generate standard security data.
[0058] The trusted calibration module is used to perform false alarm calibration and cross-source consistency calibration on standard security data, and generate trusted security data and data source trustworthiness.
[0059] The event fusion module is used to merge target records, area records, and alarm records from trusted security data to generate trusted security event data.
[0060] The relationship graph construction module is used to construct a multi-source security relationship graph based on trusted security event data, data source credibility, and security constraint data.
[0061] The risk characterization module is used to input the multi-source security relationship diagram into the improved GATNE model and generate risk characterization vectors.
[0062] The risk transmission module is used to write back the risk representation vector to the object node, event node, and region node, and generate a risk transmission relationship table containing the event time period.
[0063] The risk location module is used to correct multi-source alarm conflicts based on the risk transmission relationship table and the reliability of the data source, and generate risk location results.
[0064] The assessment output module is used to generate security risk assessment results based on risk location results and security constraint data.
[0065] The beneficial effects of this invention are:
[0066] This invention collects multi-source security data, historical false alarm data, and security constraint data, and standardizes data from different sources, formats, and time bases to generate standard security data. This allows data from video recognition, access control, perimeter alarms, fire detection, environmental sensing, and security equipment status to participate in risk assessment based on a unified data foundation. By combining historical false alarm data and cross-source consistency to generate reliable security data and data source credibility, this invention reduces the impact of false alarms from single devices, time discrepancies, and source conflicts on risk judgment, improving the reliability and stability of multi-source security data fusion.
[0067] This invention further merges target records, area records, and alarm records from trusted security data to generate trusted security event data. It then combines data source trustworthiness and security constraint data to construct a multi-source security relationship graph containing object nodes, area nodes, event nodes, trusted relationship edges, and area constraint edges, enabling a unified representation of objects, areas, events, and constraint relationships within the security system. By improving the GATNE model to embed security object attributes, calibrate trusted relationship edge embedding, perform relationship type attention fusion, and project multi-granularity security risks onto the multi-source security relationship graph, this invention enhances the ability to model heterogeneous relationships in complex security scenarios, enabling risk representation vectors to simultaneously reflect the correlation characteristics between object risks, area risks, and event risks.
[0068] This invention writes risk representation vectors back to object nodes, event nodes, and region nodes, and generates a risk transmission relationship table containing event time periods along trusted relationship edges and region constraint edges. This enables the risk assessment process not only to identify the existence of risks but also to describe the transmission relationship of risks among objects, events, and regions. By combining the risk transmission relationship table and data source credibility to correct multi-source alarm conflicts within the same object, region, and time period, and by tracing the initial event node and associated region nodes, this invention improves the accuracy and traceability of risk location results. It provides more reliable data support for risk object confirmation, risk area identification, risk level classification, and priority determination, thereby enhancing the accuracy, interpretability, and efficiency of intelligent security risk assessment. Attached Figure Description
[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0070] Figure 1 This is a flowchart of an intelligent security risk assessment method based on multi-source data fusion proposed in this invention;
[0071] Figure 2This is a schematic diagram of the structure of an improved GATNE model for an intelligent security risk assessment method based on multi-source data fusion proposed in this invention.
[0072] Figure 3 This is a schematic diagram of the structure of an intelligent security risk assessment system based on multi-source data fusion proposed in this invention. Detailed Implementation
[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0074] refer to Figure 1 and Figure 2 A smart security risk assessment method based on multi-source data fusion includes:
[0075] Collect multi-source security data, historical false alarm data, and security constraint data, and preprocess them to generate standard security data;
[0076] Based on historical false alarm data and cross-source consistency calibration standard security data, generate trusted security data and data source credibility;
[0077] Merge target records, area records, and alarm records from trusted security data to generate trusted security event data;
[0078] A multi-source security relationship graph is constructed based on trusted security event data, data source credibility, and security constraint data. The multi-source security relationship graph includes object nodes, region nodes, event nodes, trusted relationship edges, and region constraint edges.
[0079] The multi-source security relationship graph is input into the improved GATNE model to generate a risk representation vector. The improved GATNE model includes a security object attribute embedding layer, a trusted relationship edge embedding calibration layer, a relationship type attention fusion layer, a multi-granularity security risk projection layer, and a heterogeneous Skip-Gram output layer.
[0080] Write the risk representation vector back to the object node, event node, and region node, and generate a risk transmission relationship table containing the event time period along the trusted relationship edge and the region constraint edge;
[0081] Based on the risk transmission relationship table and the credibility of the data source, the conflict of multiple alarms from the same object, the same region, and the same time period is corrected, the first event node and related regional nodes are traced, and the risk location results are generated.
[0082] Based on the risk location results and security constraint data, a security risk assessment result is generated, which includes the risk object, risk area, risk type, risk level, and handling priority.
[0083] In this embodiment, the preprocessing to generate standard security data includes:
[0084] The system identifies data items for multi-source security data, historical false alarm data, and security constraint data, and performs source identifier completion, field name unification, data format conversion, and anomaly sampling and removal. The multi-source security data includes video recognition data, access control data, perimeter alarm data, fire sensing data, environmental sensor data, and security device status data. Historical false alarm data includes false alarm event identifiers, false alarm source devices, false alarm occurrence areas, false alarm occurrence times, false alarm trigger types, and false alarm handling results. Security constraint data includes area sensitivity levels, access permission constraints, deployment time constraints, device coverage constraints, and risk handling level constraints. The area sensitivity levels are derived from a preset area configuration table on the security platform, divided into Level 1, Level 2, Level 3, and Level 4 according to area usage and control requirements. Level 1 corresponds to ordinary public areas, Level 2 to key patrol areas, Level 3 to restricted access areas, and Level 4 to core control areas. Access permission constraints, deployment time constraints, device coverage constraints, and risk handling level constraints are all stored in association with area numbers.
[0085] Based on the data collection time, device timestamp, alarm trigger time, and response time, and by unifying the time benchmark and segmenting the event time period, time-series security data is generated. Specifically, the generation of time-series security data involves unifying the time benchmark and segmenting the event time period.
[0086] Using the security platform's time synchronization as a unified time benchmark, the collection time, device timestamp, alarm trigger time, and handling time are converted into millisecond-level timestamps. The difference between the device timestamp and the collection time is calculated based on the device number. The median of the differences of 30 consecutive records for the same device is taken as the time offset. Device timestamps with an absolute offset greater than 2 seconds are corrected according to the time offset. The alarm trigger time is used as the event start time, and the handling time is used as the event end time. If the handling time is missing, 60 seconds after the event start time is used as the event end time. Event time periods are divided into fixed 30-second windows. Cross-window events are split according to the window boundaries. The window start and end times, object identifier, area number, device number, event type, and source time fields are recorded to generate time-series security data.
[0087] Based on object identifiers, area numbers, device numbers, event types, and constraint types, time-series security data, historical false alarm data, and security constraint data are associated, missing field imputation, duplicate record merging, and sensitive field desensitization are performed to generate standard security data. Specifically, the missing field imputation, duplicate record merging, and sensitive field desensitization processes for generating standard security data are as follows:
[0088] Integrity markers are set for object identifiers, area numbers, device numbers, event types, constraint types, and event time periods. When an object identifier is missing, a unique target record is matched using the device number, area number, event type, and event time period to fill it in. When an area number is missing, the coverage area is matched using the device coverage constraint to fill it in. When an event type and constraint type are missing, the mapping relationship between alarm type and constraint type is used to fill it in. Using object identifiers, area numbers, device numbers, event types, and event time periods as merging keys, records with the same merging key or alarm trigger intervals not exceeding 1 second are merged. The earliest collection time is taken, and the latest processing time is taken. The source device is deduplicated and retained. The fields of personnel name, ID number, contact number, vehicle license plate, and precise location are anonymized. The personnel name is replaced with the personnel category, the ID number, contact number, and vehicle license plate are hashed using a fixed anonymized salt value, and the precise location is replaced with the area number to form standard security data.
[0089] In this embodiment, the generation of trusted security data and the trustworthiness of the data source include:
[0090] Based on historical false alarm data, the number of false alarms, the number of valid alarms, and the number of times the handling results were consistent are statistically analyzed according to the source device, event type, occurrence area, and occurrence time. A historical false alarm characteristic table is then generated. Specifically, the generation of the historical false alarm characteristic table is as follows:
[0091] The statistical key is defined by the 30-minute time period corresponding to the false alarm source device, false alarm trigger type, false alarm occurrence area, and false alarm occurrence time. The 30-minute time period is divided sequentially starting from 00:00 each day. The handling result field in the historical false alarm data is traversed. When the handling result is a false alarm, no abnormality, or repeated alarm, the false alarm count is incremented by 1. When the handling result is a valid risk, confirmed abnormality, or handled, the valid alarm count is incremented by 1. The handling result that appears most frequently within the same statistical key is used as the baseline handling result. When the number of occurrences is equal, the baseline handling result is determined in the order of valid risk, confirmed abnormality, handled, false alarm, no abnormality, and repeated alarm. Records with handling results consistent with the baseline handling result are counted in the handling result consistency count. The statistical key, the number of false alarms, the number of valid alarms, and the number of handling result consistency counts are combined to generate a historical false alarm feature table.
[0092] Event matching is performed on video recognition records, access control records, perimeter alarm records, and environmental perception records within the same target, area, and time period in standard security data to generate cross-source consistent records. Specifically, the event matching for video recognition records, access control records, perimeter alarm records, and environmental perception records within the same target, area, and time period in standard security data is as follows:
[0093] A matching primary key is established using object identifier, area number, and 30-second event time period number. The target identifier and recognition time of video recognition records, the access object and access time of access control records, the alarm point and trigger time of perimeter alarm records, and the sensor point and sampling time of environmental perception records are extracted. Records that fall within the same event time period and have the same area number are written into the same candidate event group. When the object identifier of the video recognition record and the access control record are the same, the target consistency flag is set to 1. When the points of the perimeter alarm record and the environmental perception record fall within the same device coverage area constraint, the area consistency flag is set to 1. When the candidate event group contains records from at least two sources and the target consistency flag or the area consistency flag is set to 1, it is determined to be a cross-source consistent record.
[0094] The data source credibility is obtained by weighted summation of historical false alarm suppression score, cross-source consistency score, complete data collection score, and time synchronization score. Based on the data source credibility, standard security data is then subjected to credibility marking, conflict record downweighting, and low-credibility record removal to generate credible security data. Specifically, the data source credibility is obtained by weighted summation of historical false alarm suppression score, cross-source consistency score, complete data collection score, and time synchronization score.
[0095] Using the source device as the scoring object, the number of false alarms in the historical false alarm feature table is divided by the sum of the number of false alarms and the number of valid alarms to obtain the historical false alarm ratio. When the denominator is 0, the historical false alarm ratio is 0.5. The historical false alarm suppression score is 1 minus the historical false alarm ratio. The number of cross-source consistent records is divided by the number of candidate event groups in which the same source device participates to obtain the cross-source consistency score. When the denominator is 0, the score is 0.5. The number of non-empty required fields is divided by the total number of required fields to obtain the completeness score. Required fields include object identifier, region number, device number, event type, event time period, and source time fields. The absolute value of the device time offset is divided by 2000 milliseconds to obtain the synchronization deviation ratio. The time synchronization score is 1 minus the synchronization deviation ratio, with a minimum value of 0. The four scores are weighted and summed in order of weights of 0.35, 0.30, 0.20, and 0.15 to obtain the data source credibility.
[0096] The process involves performing trusted labeling, conflict record downgrading, and low-trust record removal on standard security data based on the trustworthiness of the data source, thereby generating trusted security data. Specifically:
[0097] Write the data source credibility into the standard security data record corresponding to the source device; record with a data source credibility of not less than 0.80 is marked as a trusted record, record with a credibility of greater than or equal to 0.50 and less than 0.80 is marked as a record to be verified, and record with a credibility of less than 0.50 is discarded as a low-credibility record; establish a conflict group for records with inconsistent event types, area numbers, and event time periods, or conflicting access status and alarm status, retain the record with the highest data source credibility in the group, and change the record weight of the remaining records in the group to the data source credibility multiplied by 0.5; combine trusted records, records to be verified, record weights, and conflict group numbers to generate trusted security data.
[0098] In this embodiment, generating trusted security event data includes:
[0099] Candidate security event records are generated by combining target records, area records, and alarm records according to target identifier, area number, device number, alarm type, and event time period.
[0100] The event merge score is obtained by subtracting the time interval penalty value from the weighted result of target similarity, region overlap, alarm type consistency, and data source credibility. Specifically, the event merge score is as follows:
[0101] Using two candidate security event records as the calculation unit, the target similarity is set to 1 if the target identifiers are the same, 0.8 if the target identifiers are different but the de-identified target codes are the same, and 0 if neither condition is met. The overlap area of the two regions is divided by the union area to obtain the region overlap, which is set to 1 if the region number is the same and there is no coverage area data. The alarm type consistency is set to 1 if the alarm type belongs to the same constraint type in the security constraint data, and 0 otherwise. The average of the data source credibility of the two records is taken as the credibility score. The target similarity, region overlap, alarm type consistency, and credibility score are multiplied by 0.35, 0.25, 0.25, and 0.15 respectively and then summed. The absolute value of the time difference between the start times of the two events is divided by 30000 milliseconds and then multiplied by 0.2 to obtain the time interval penalty value, which is set to 0.2 if the penalty value is greater than 0.2. The event merge score is obtained by subtracting the time interval penalty value from the weighted result.
[0102] Candidate security event records that meet the merging threshold in the event merging score are grouped into a single trusted security event. This involves merging target records, area records, alarm records, source devices, and event time periods to generate trusted security event data. Specifically, the generation of trusted security event data includes:
[0103] Using an event merging score of no less than 0.75 as the merging threshold, candidate security event records that meet the merging threshold are established into merging relationships. Candidate security event records with direct or indirect merging relationships are grouped into the same event group. Within the same event group, target records are deduplicated by object identifier, area records by area number, alarm records by alarm type and alarm trigger time, and source devices by device number. The event start time is the earliest alarm trigger time within the event group, and the event end time is the latest handling time within the event group. If the handling time is missing, 60 seconds after the latest alarm trigger time is used. The event group number, target record, area record, alarm record, source device, event start time, event end time, and highest data source credibility are written into the event data table to generate trusted security event data.
[0104] In this embodiment, the construction of a multi-source security relationship diagram based on trusted security event data, data source credibility, and security constraint data includes:
[0105] Map target records, area records, and alarm records to object nodes, area nodes, and event nodes, configure node types, node sources, and event time periods, and generate a security node table. Specifically, generating the security node table involves:
[0106] Using the event group number in the trusted security event data as the node affiliation field, the object identifier, object category, source device, and event time period in the target record are mapped to object nodes; the area number, area sensitivity level, device coverage, and event time period in the area record are mapped to area nodes; and the alarm type, alarm trigger time, handling result, source device, and event time period in the alarm record are mapped to event nodes. A unique node identifier is generated according to the node type, object identifier, area number, alarm type, alarm trigger time, and event group number, and the node source and data source trustworthiness are configured. Node records with the same unique node identifier are deduplicated, and the node attribute with the highest data source trustworthiness is retained to generate a security node table.
[0107] Matching the relationships between object nodes, region nodes, and event nodes, and generating trusted relationship edges, specifically, the generation of trusted relationship edges involves:
[0108] Using event group number, object identifier, region number, and event time period as matching fields, the system iterates through object nodes, region nodes, and event nodes in the security node table. When the object identifier of an object node matches the target record to which the event node belongs, a trusted object-event relationship edge is established. When the source device of an event node falls within the device coverage area of a region node, or when the region number of an event node matches that of a region node, a trusted event-region relationship edge is established. When an object node and a region node belong to the same event group number and have the same region number, a trusted object-region relationship edge is established. Edge trustworthiness is determined by the data source trustworthiness and the event time sequence. The adjacency and regional location matching degree are weighted and obtained. The data source credibility is the average of the data source credibility of the two nodes at both ends of the edge. The event time adjacency is 1 minus the ratio of the absolute value of the time difference between the start time of the events of the two nodes to 30000 milliseconds, with a minimum value of 0. The regional location matching degree is 1 when the region number is the same, 0.8 when the device coverage areas intersect, and 0 when neither condition is met. The three terms are multiplied by 0.5, 0.3 and 0.2 respectively and then summed. Edge credibility is retained as a credible relation edge when it is not less than 0.60. A credible relation edge containing the start node, end node, edge type, edge credibility and event time period is generated.
[0109] Generate regional constraint edges based on regional sensitivity level, access permission constraints, deployment time constraints, and equipment coverage constraints. Specifically, the generation of regional constraint edges involves:
[0110] Match the area number in the security constraint data with the area node in the security node table; when the area sensitivity level is level three or four, establish an area sensitivity constraint edge between the area node and the event node; when the object identifier corresponding to the object node is not included in the authorized object list, establish an access permission constraint edge between the object node and the area node; when the alarm trigger time of the event node falls within the closed interval of the arming start and end time, establish an arming time constraint edge between the event node and the area node; when the source device number of the event node is included in the device coverage area, establish an device coverage constraint edge between the event node and the area node; write the edge type, start node, end node, constraint label, area sensitivity level, and effective time period into the area constraint edge record to generate the area constraint edge;
[0111] Encapsulate object nodes, region nodes, event nodes, trusted relationship edges, and region constraint edges to generate a multi-source security relationship graph.
[0112] In this embodiment, the generation of the risk representation vector includes:
[0113] Based on the GATNE model's basic embedding, edge embedding, global node embedding generation, and heterogeneous Skip-Gram output, an improved GATNE model is formed by adding a security object attribute embedding layer at the basic embedding input, a trusted relationship edge embedding calibration layer at the edge embedding branch, a relationship type attention fusion layer at the edge type self-attention combination position, and a multi-granularity security risk projection layer at the global node embedding generation output. The training objects of the heterogeneous Skip-Gram output layer are set as object nodes, region nodes, and event nodes.
[0114] This invention structurally optimizes the GATNE model and proposes an improved GATNE model for intelligent security risk assessment scenarios involving multi-source data fusion. The model is based on basic embedding, edge embedding, global node embedding generation, and heterogeneous Skip-Gram output. It supplements node type, node source, event time period, and node attribute information through a security object attribute embedding layer, so that object nodes, region nodes, and event nodes have a unified attribute expression basis.
[0115] In the edge embedding process, a trusted relation edge embedding calibration layer is introduced, which writes the edge trustworthiness and constraint label into the edge representation, so that trusted relation edges and regional constraint edges can carry source reliability and security constraint information to participate in model calculation, reducing the interference of low-trust alarms and conflict records on risk representation.
[0116] In the relationship fusion stage, the relationship type attention fusion layer performs differentiated fusion of object-event relationships, event-region relationships, object-region relationships, and region constraint relationships, enabling the model to highlight key relationships related to risk transmission; the multi-granularity security risk projection layer further divides the fusion results into object-granularity, region-granularity, and event-granularity risk embeddings, improving the risk representation vector's ability to distinguish risk objects, risk regions, and risk events;
[0117] The heterogeneous Skip-Gram output layer uses object nodes, region nodes, and event nodes as training objects, so that the node context prediction results are consistent with the security risk association structure, and enhance the model's ability to express the risk transmission path and the relationship of the first event tracing.
[0118] The improved GATNE model, while retaining the ability to embed heterogeneous graphs, enhances the ability to express node attributes, calibrate edge credibility, distinguish relationship types, and characterize risks at multiple granularities, making the generated risk representation vectors more suitable for multi-source alarm conflict correction, risk transmission analysis, and security risk assessment.
[0119] The multi-source security relationship diagram is input into the security object attribute embedding layer, and a security node embedding matrix is generated according to node type, node source, event time period, and node attribute. The security object attribute embedding layer includes:
[0120] Node basic coding unit: Reads the node type, node source and event time period from the security node table, converts object nodes, area nodes and event nodes into node type codes, converts source device and data source trustworthiness into node source codes, and converts event time period into time location codes;
[0121] Node attribute encoding unit: Reads the object identifier and object category in the object node, the area number and area sensitivity level in the area node, and the alarm type and handling result in the event node, converts discrete fields into attribute codes, and converts area sensitivity levels into level numerical codes;
[0122] Embedded fusion output unit: According to the node identifier, the node type encoding, node source encoding, time location encoding and attribute encoding are concatenated, linearly mapped, weighted and normalized to form a security node embedding matrix;
[0123] In the security object attribute embedding layer, the node basic coding unit reads the node type, node source, and event time period according to the node arrangement order in the security node table. The node type uses three fixed codes: object node codes are 1, 0, 0; area node codes are 0, 1, 0; and event node codes are 0, 0, 1. The node source generates a source device code based on the source device number, and uses the data source credibility as a product coefficient to form the node source code. The event time period generates a time location code based on the 30-second event time period number. The node attribute coding unit reads the node attribute fields according to the node type, performs discrete number mapping on the object identifier, object category, area number, alarm type, and handling result, and converts the area sensitivity level from level 1 to level 4 into a level numerical code of 0.25, 0.50, 0.75, and 1.00. The embedded fusion output unit is concatenated into a node concatenation vector in a fixed order of node type encoding, node source encoding, time location encoding, and attribute encoding. The node concatenation vector is multiplied by a trainable weight matrix and a bias vector is added to obtain an initial node embedding vector. The initial node embedding vector is multiplied by the corresponding data source credibility to obtain a credible node embedding vector. Each component in the credible node embedding vector is divided by the square root of the sum of the squares of the vectors to form a normalized node embedding vector. All normalized node embedding vectors are arranged according to node identifiers to generate a security node embedding matrix.
[0124] The security node embedding matrix, trusted relationship edges, region constraint edges, and edge trustworthiness are input into the trusted relationship edge embedding calibration layer. A trusted edge embedding matrix is generated according to edge type, edge connecting node, edge trustworthiness, and constraint label. The trusted relationship edge embedding calibration layer includes:
[0125] Edge Record Organization Unit: Consolidates trusted relationship edges and regional constraint edges into an edge record table, retaining edge type, start node, end node, edge trustworthiness, constraint label, and event time period;
[0126] Edge connection coding unit: Based on the start node and end node in the edge record table, extract the corresponding node embedding vector from the security node embedding matrix to form an edge connection code with connection direction;
[0127] Trusted constraint calibration unit: performs trusted weighting on edge connection encoding based on edge trustworthiness, and converts constraint labels into constraint label encoding to form a trusted constraint edge representation;
[0128] Edge embedding output unit: Combines edge type encoding, trusted constraint edge representation and event time period encoding in a fixed order to generate a trusted edge embedding matrix;
[0129] In the trusted relation edge embedding calibration layer, the edge recording and organization unit establishes an edge record table according to the generation order of trusted relation edges and regional constraint edges. Trusted relation edges for object events, trusted relation edges for event regions, trusted relation edges for object regions, sensitive constraint edges for regions, access permission constraint edges, deployment time period constraint edges, and device coverage constraint edges are written into their corresponding edge types. The edge connection encoding unit retrieves the security node embedding matrix according to the start node and end node, obtaining the start node embedding vector and the end node embedding vector. The start node embedding vector, the end node embedding vector, and the direction difference vector obtained by subtracting the start node embedding vector from the end node embedding vector are then concatenated sequentially to form the edge connection code. The trusted constraint calibration unit multiplies each component in the edge connection code by the edge trustworthiness, and the regional constraint... When an edge lacks a credibility value, it is set to 1. The area sensitivity constraint, access permission constraint, deployment time period constraint, and equipment coverage constraint are converted into different constraint label codes. When a trusted relation edge has no constraint label, it is written into the unconstrained label code. The constraint label code is then concatenated to the trusted weighted edge connection code to form a trusted constraint edge representation. The edge embedding output unit concatenates the edge type code, trusted constraint edge representation, and event time period code in a fixed order to form an edge input vector. The edge input vector is multiplied by the trained linear mapping matrix and a bias vector is added to obtain the initial edge embedding vector. The initial edge embedding vector is then normalized by dividing it by the vector magnitude. When the vector magnitude is 0, it is set to 1. All normalized edge embedding vectors are combined according to the edge arrangement order in the edge record table to generate a trusted edge embedding matrix.
[0130] The security node embedding matrix and the trusted edge embedding matrix are input into the relation type attention fusion layer. A relation fusion node embedding matrix is generated according to the relation type of trusted relation edges and region constraint edges. The relation type attention fusion layer includes:
[0131] Relationship type grouping unit: Read the edge type, start node, end node and event time period in trusted relationship edges and regional constraint edges, and classify the trusted relationship edges of object events, trusted relationship edges of event regions, trusted relationship edges of object regions, sensitive constraint edges of regions, access permission constraint edges, deployment time period constraint edges and device coverage constraint edges into the corresponding relationship type groups, and establish a set of node association edges with node identifier as index;
[0132] Relational adjacency construction unit: According to the starting node and ending node in the node association edge set, extract the center node embedding vector and the association node embedding vector from the security node embedding matrix, extract the corresponding trusted edge embedding vector from the trusted edge embedding matrix, and concatenate the association node embedding vector and the trusted edge embedding vector in a fixed order to form a relational adjacency representation;
[0133] Relationship scoring normalization unit: Calculates node similarity score, edge credibility score, constraint matching score, and time period adjacency score, and generates edge relationship score according to fixed weight; averages the edge relationship scores within the same relationship type group to obtain the relationship type score; normalizes the relationship type scores corresponding to the same node to obtain the relationship type weight;
[0134] Relation fusion output unit: Based on the edge relationship score and relationship type weight, the relation adjacency representation is weighted and aggregated at two levels to generate a relation fusion node embedding matrix;
[0135] In the relation type attention fusion layer, the relation type grouping unit reads the node identifiers according to the node arrangement order in the security node table, writes the trusted relation edges and region constraint edges connected to the current node into the node association edge set, and forms relation type groups according to edge type. The relation adjacency construction unit takes the current node as the center node and the other end node of the edge as the associated node, extracts the center node embedding vector, the associated node embedding vector, and the trusted edge embedding vector, and concatenates the associated node embedding vector and the trusted edge embedding vector to form the relation adjacency representation. The relation scoring normalization unit divides the dot product of the embedding vector of the central node and the embedding vector of the associated node by the product of their magnitudes. When the magnitude is 0, the node similarity score is 0. The result is then added by 1 and divided by 2 to obtain the node similarity score in the range of 0 to 1. The edge credibility score is taken from the edge credibility in the edge record. When the edge credibility of the region constraint edge is missing, it is taken as 1. When the constraint label is empty, the constraint matching score is 0.5. When the constraint label is not empty, the constraint matching score is 1. The score for adjacent events in the same time period is 1. The score for event periods with a difference of 1 30 seconds is 0.5. The score for event periods with a difference of at least 2 30 seconds is 0. The node similarity score, edge credibility score, constraint matching score, and time period adjacent score are multiplied by 0.35, 0.30, 0.20, and 0.15 respectively and then added to obtain the edge relation score. The average of all edge relation scores within the same relation type group is used to obtain the relation type score. The relation type score for each relation type corresponding to the same node is divided by the sum of all relation type scores to obtain the relation type weight. When the sum of all relation type scores is 0, the weights are evenly distributed according to the number of relation types. Within the same relation type group, the relation fusion output unit multiplies each relation adjacency representation by the corresponding edge relation score, then divides it by the sum of all edge relation scores in the same group to obtain the relation type adjacency vector. When the sum of all edge relation scores in the same group is 0, the vectors are evenly distributed according to the number of edges. The adjacency vectors of each relation type are multiplied by the corresponding relation type weight and then summed to obtain the node relation fusion vector. The embedding vector of the central node is multiplied by 0.4 and the node relation fusion vector is multiplied by 0.6 and then added together. The sum is then divided by the vector magnitude for normalization. All normalized results are combined according to the node arrangement order in the security node table to generate the relation fusion node embedding matrix.
[0136] The relationship fusion node embedding matrix is input into the multi-granularity security risk projection layer, generating a multi-granularity risk embedding matrix at the object granularity, region granularity, and event granularity. The multi-granularity security risk projection layer includes:
[0137] Granularity node partitioning unit: Read the node types in the security node table and divide the node embedding vector in the relationship fusion node embedding matrix into object granularity node vector, region granularity node vector and event granularity node vector;
[0138] Granularity-based risk projection unit: Performs linear projection and activation processing on object granularity node vectors, region granularity node vectors and event granularity node vectors respectively to generate object risk embedding, region risk embedding and event risk embedding;
[0139] Granularity Consistency Calibration Unit: Based on the connection relationship between trusted relation edges and regional constraint edges, it performs consistency calibration on the risk embedding of connected objects, regional risk embeddings, and event risk embeddings within the same event time period;
[0140] Risk embedding output unit: Combines the calibrated risk embeddings according to the order of object granularity, region granularity, and event granularity to generate a multi-granularity risk embedding matrix;
[0141] In the multi-granularity security risk projection layer, the granularity node partitioning unit reads the node type according to the node arrangement order in the security node table. When the node type is an object node, the corresponding relationship fusion node embedding vector is written into the object granularity node vector set; when the node type is a region node, it is written into the region granularity node vector set; and when the node type is an event node, it is written into the event granularity node vector set. The granular risk projection unit multiplies each object granularity node vector, region granularity node vector, and event granularity node vector by the object projection matrix, region projection matrix, and event projection matrix, respectively, and adds the object projection bias, region projection bias, and event projection bias, respectively. The projection matrix and projection bias are updated during the training phase of the improved GATNE model. Components less than 0 in the projection result are set to 0, while components not less than 0 retain their original values, forming the initial risk embedding for the corresponding granularity. The granularity-consistent calibration unit reads the start node, end node, and event time period of the trusted relation edges and regional constraint edges. It groups connected object nodes, regional nodes, and event nodes within the same event time period into calibration groups and calculates the average vector of each initial risk embedding within the calibration group. Each initial risk embedding is multiplied by 0.7, and the average vector is multiplied by 0.3 before being summed to obtain the calibration risk embedding. The risk embedding output unit divides each calibration risk embedding by the vector magnitude, setting it to 1 when the vector magnitude is 0, to obtain the normalized risk embedding. All normalized risk embeddings are arranged sequentially by object granularity, regional granularity, and event granularity to generate a multi-granularity risk embedding matrix.
[0142] The multi-granularity risk embedding matrix is input into the heterogeneous Skip-Gram output layer. Context node prediction results are generated based on the node walk sequence, and the multi-granularity risk embedding matrix is aggregated to generate a risk representation vector. The heterogeneous Skip-Gram output layer includes:
[0143] Node traversal construction unit: Starting from object nodes, region nodes, and event nodes, a node traversal sequence is formed based on trusted relation edges and region constraint edges, and the center node, traversed nodes, relation types, and event time periods are recorded;
[0144] Context sample extraction unit: Extracts the central node and context node from the node walk sequence according to a fixed context window to form context node samples;
[0145] Context node prediction unit: Calculates the prediction probability between the center node and the context node based on the multi-granularity risk embedding matrix, and generates the context node prediction result;
[0146] Risk representation aggregation unit: Based on the prediction results of context nodes, the multi-granularity risk embedding matrix is aggregated with probability weighting to generate a risk representation vector;
[0147] In the heterogeneous Skip-Gram output layer, the node traversal construction unit reads object nodes, region nodes, and event nodes according to the node arrangement order in the security node table, using each node as the starting point for a traversal. Each traversal is 6 nodes long. When a current node has multiple optional connection edges, the edge selection score for trusted relationship edges is the edge's confidence level, while the edge selection score for region constraint edges is 1. The connection edge with the highest selection score is selected to proceed to the next node. If edge selection scores are the same, selection is based on the event time period from nearest to farthest, forming a node traversal sequence. The context sample extraction unit uses each node in the node traversal sequence as the center node, and takes two nodes before and after the center node as context nodes. Positions outside the sequence boundary are not assigned values, forming a context node sample containing the center node identifier, context node identifier, relationship type, and event time period. The context node prediction unit extracts the center node risk embedding vector and the context node risk embedding vector from the multi-granularity risk embedding matrix. It divides the dot product of the two vectors by the product of their magnitudes; when the magnitude is 0, the prediction score is 0. It then takes the exponent of the prediction scores for each context node corresponding to the same center node and divides each exponent by the sum of all prediction score exponents to obtain the context node prediction probability. The context node identifier and the context node prediction probability are combined to generate the context node prediction result. The risk representation aggregation unit multiplies the context node risk embedding vectors corresponding to the same center node by the context node prediction probability and sums them to obtain the context risk aggregation vector. It then multiplies the center node risk embedding vector by 0.6 and the context risk aggregation vector by 0.4 and adds them together to obtain the node risk representation vector. Finally, it normalizes the node risk representation vector by dividing it by the vector magnitude; when the magnitude is 0, it is set to 1. All normalized node risk representation vectors are arranged by node identifier to generate the risk representation vector.
[0148] An improved GATNE model is trained using a weighted sum of context node prediction error, credible relation edge reconstruction error, and multi-granularity risk embedding consistency error. This trained improved GATNE model is then used to generate risk representation vectors. The training of the improved GATNE model specifically involves:
[0149] Using a multi-source security relationship graph, node traversal sequence, trusted relationship edges, regional constraint edges, and a multi-granularity risk embedding matrix as training inputs, the context node prediction error is calculated using the context node prediction results. When the prediction probability of the center node corresponding to the real context node is less than 0.000001, it is set to 0.000001. The natural logarithm of the real context node prediction probability is taken and negative. After traversing all context node samples, the average is calculated to obtain the context node prediction error. The trusted relationship edge reconstruction error is calculated using trusted relationship edges. The starting and ending nodes in the trusted relationship edge are used as positive samples, labeled 1. The same number of negative samples are extracted from node pairs that are not connected by trusted relationship edges within the same event period, labeled 0. The risk embeddings at both ends of the node pair and the corresponding trusted edge embeddings are concatenated and input into the edge reconstruction mapping parameters to obtain the edge reconstruction probability in the range of 0 to 1. The positive sample error is the negative value of the natural logarithm of the edge reconstruction probability, and the negative sample error is the negative value of the natural logarithm of 1 minus the edge reconstruction probability. The average of all positive sample errors and negative sample errors is calculated to obtain the trusted relationship edge reconstruction error. The multi-granularity risk embedding consistency error is calculated for object nodes, region nodes, and event nodes connected by trusted relation edges and region constraint edges. The group average vector is obtained by averaging the risk embedding vectors within the same connection group according to their components. The group average vector is then obtained by subtracting each risk embedding vector from the group average vector according to its components, squaring the results, summing them, and averaging the sum over the connection group. The training optimization target is then obtained by multiplying the context node prediction error by 0.50, the trusted relation edge reconstruction error by 0.30, and the multi-granularity risk embedding consistency error by 0.20. The training parameters in the security object attribute embedding layer, the trustworthy relation edge embedding calibration layer, the relation type attention fusion layer, the multi-granularity security risk projection layer, and the heterogeneous Skip-Gram output layer are updated with gradients according to the training optimization objective. The learning rate is 0.001. Each training round traverses all node walk sequences and edge samples. Training is stopped when the training rounds reach 100 rounds or the optimization objective decreases by less than 0.0001 for 5 consecutive rounds. The model parameters with the smallest training optimization objective are saved to obtain the improved GATNE model after training is completed.
[0150] In this embodiment, generating a risk transmission relationship table containing event time periods along the trusted relationship edge and the region constraint edge includes:
[0151] The risk representation vector is written back to the object node, event node, and region node according to the node identifier, generating a node risk write-back table. Specifically, the generation of the node risk write-back table is as follows:
[0152] The system reads the node identifier, node type, node source, event time period, and node attributes from the security node table, and reads the node risk representation vector corresponding to the node identifier from the risk representation vector. Using the node identifier as the matching key, it searches the node records in the security node table one by one, writes the node risk representation vector with the same node identifier to the corresponding node record, and writes a vector of all zeros and marks it as missing for write-back when no node risk representation vector is found. It calculates the magnitude of the node risk representation vector by squaring each component of the node risk representation vector, summing them, and taking the square root to obtain the magnitude. When the magnitude is 0, the risk intensity value is 0. When the magnitude is not 0, the magnitude is divided by the maximum magnitude of all nodes to obtain the risk intensity value in the range of 0 to 1. The system writes the node identifier, node type, node source, event time period, node attributes, node risk representation vector, risk intensity value, and write-back mark into a table according to the node arrangement order in the security node table to generate a node risk write-back table.
[0153] Traverse trusted relation edges and region constraint edges, combine edge connection direction, edge type, edge trustworthiness, and constraint label to generate candidate risk transmission records. Specifically, generating candidate risk transmission records involves:
[0154] Read the starting node, ending node, edge type, edge credibility, constraint label, and event time period in the order of edge records for trusted relation edges and regional constraint edges; use the starting node and ending node as matching keys to retrieve the corresponding node risk representation vector and risk intensity value in the node risk write-back table; discard the corresponding edge record if no node is found; use the starting node pointing to the ending node as the edge connection direction; if the event time period of the starting node is later than that of the ending node, swap the starting node and ending node so that the edge connection direction points from the earlier event time period to the later event time period; if the event time periods are the same, retain the original edge connection direction. The edge credibility of a trusted relation edge is taken from the edge credibility in the edge record. The edge credibility of a region constraint edge is 1 when the edge credibility field is empty. When the constraint label is empty, an unconstrained label is written. When the constraint label is not empty, the corresponding label of the region sensitive constraint, access permission constraint, deployment time period constraint or device coverage constraint is retained. The starting node, ending node, edge connection direction, edge type, edge credibility, constraint label, event time period, starting node risk representation vector, ending node risk representation vector, starting node risk intensity value and ending node risk intensity value are written into the candidate record, and the candidate risk transmission record is generated in the order of the edge record.
[0155] Calculate the risk transmission score of candidate risk transmission records. The risk transmission score is obtained by subtracting the event time interval penalty value from the weighted average of the risk representation similarity, edge credibility, and constraint matching value between the starting and ending nodes. Specifically, the calculation of the risk transmission score of candidate risk transmission records is as follows:
[0156] Read the risk representation vectors of the starting and ending nodes, edge credibility, constraint labels, and event periods of the starting and ending nodes from each candidate risk transmission record. Multiply the risk representation vectors of the starting and ending nodes by their corresponding components and sum them to obtain a vector dot product. Sum the squares of the two risk representation vectors by their respective components and take the square root to obtain the magnitudes of the two vectors. Divide the vector dot product by the product of the two magnitudes to obtain a cosine similarity value. The cosine similarity value is 0 when any vector magnitude is 0. Add 1 to the cosine similarity value and divide by 2 to obtain the risk representation similarity in the range of 0 to 1. Edge credibility is directly taken from the edge credibility in the candidate risk transmission record; it is 0 when less than 0 and 1 when greater than 1. When the constraint label is unconstrained, the constraint... The matching value is set to 0.50. When the constraint label is a regional sensitive constraint, the regional sensitivity level is 0.75 for level 3 and 1.00 for level 4. When the constraint label is a access permission constraint, deployment time constraint, or device coverage constraint, the constraint matching value is set to 1.00. Calculate the number of 30-second event periods between the start node event period and the end node event period. Multiply the difference by 0.05 to obtain the event period interval penalty value. If the event period interval penalty value is greater than 0.20, set it to 0.20. Multiply the risk representation similarity by 0.45, the edge credibility by 0.35, and the constraint matching value by 0.20, sum them, and then subtract the event period interval penalty value to obtain the risk transmission score. If the risk transmission score is less than 0, set it to 0. Write the risk transmission score into the corresponding candidate risk transmission record.
[0157] Candidate risk transmission records that meet the transmission threshold are screened, and risk transmission relationship tables containing the event time period are generated by organizing them according to the event time period. Specifically, the risk transmission relationship table is generated by organizing the risk transmission relationship according to the event time period as follows:
[0158] Read the starting node, ending node, edge connection direction, edge type, edge credibility, constraint label, event time period, starting node risk intensity value, ending node risk intensity value, and risk transmission score from the candidate risk transmission records. Set the transmission threshold to 0.65, retaining candidate risk transmission records with a risk transmission score not less than 0.65 and deleting those with a score less than 0.65. Sort the retained records by event time period as the first sorting field, risk transmission score in descending order as the second sorting field, and edge credibility in descending order as the third sorting field. Remove duplicates from the retained records where the starting node, ending node, edge type, and event time period are identical, retaining the record with the highest risk transmission score. Multiply the risk transmission score by 0.50, the starting node risk intensity value by 0.25, and the ending node risk intensity value by 0.25, then add them together to obtain the transmission intensity value. Create a table with event time periods, starting node, ending node, edge connection direction, edge type, edge credibility, constraint label, risk transmission score, and transmission intensity value as table fields, and write them into the retained records in event time period order to generate a risk transmission relationship table containing event time periods.
[0159] In this embodiment, the generation of risk location results includes:
[0160] Multi-source alarm records in the risk transmission relationship table are aggregated by object node, region node, and event time period, and alarm conflict records within the same object node, the same region node, and the same event time period are marked.
[0161] By combining the data source credibility and risk transmission relationship table, a trusted correction is performed on the alarm conflict records to generate corrected alarm records. Specifically, the trusted correction of the alarm conflict records by combining the data source credibility and risk transmission relationship table involves:
[0162] Read each alarm conflict record sequentially, including the object node, region node, event node, alarm type, source device, event time period, and original alarm status. Search for the data source credibility using the source device as the matching key; if no record is found, the data source credibility is set to 0.50. Search the risk transmission relationship table for risk transmission records containing the same object node, region node, or event node with the same event time period, and take the highest risk transmission score from the search results as the transmission support value; if no record is found, the transmission support value is set to 0. Set the alarm type matching value to 1 if the alarm type matches the event type corresponding to the event node, and 0 if they do not. Multiply the data source credibility by 0.40 and the transmission support value by 0. 40. Multiply the alarm type matching value by 0.20 and add them together to obtain the alarm reliability correction score. When there are multiple alarm conflict records in the same object node, the same area node, and the same event time period, retain the alarm record with the highest alarm reliability correction score as the master correction record, and mark the remaining alarm records as subordinate conflict records. When the alarm reliability correction score of the master correction record is not less than 0.70, the corrected alarm status is taken as a valid alarm; when it is less than 0.70, it is taken as an alarm pending review. Write the object node, area node, event node, alarm type, source device, event time period, data source reliability, propagation support value, alarm reliability correction score, and corrected alarm status into a table to generate a corrected alarm record.
[0163] The event nodes corresponding to the alarm records are traced back along the trusted relationship edges and the region constraint edges to determine the initial event node. Specifically, the event nodes corresponding to the alarm records are traced back along the trusted relationship edges and the region constraint edges as follows:
[0164] Read the event node, object node, region node, event time period, data source credibility, and alarm credibility correction score from the correction alarm record, and use the event node corresponding to the correction alarm record as the backtracking starting point; search the risk transmission relationship table containing event time periods for risk transmission records where the termination node is equal to the backtracking starting point and the event time period of the starting node is not later than the event time period of the termination node, and retain records whose edge type belongs to the credible relationship edge or region constraint edge as backtracking candidate records; continue searching forward from the starting node in the backtracking candidate records, with a backtracking depth of 5 levels and a backtracking time range of 10 30-second event time periods. Records exceeding the backtracking depth or backtracking time range are not written into the backtracking path; for each backtracking path The effective score for backtracking is calculated by multiplying the average risk transmission score within the path by 0.50, the average edge credibility score within the path by 0.30, and the data source credibility score of the corrected alarm record by 0.20, and then adding them together. Backtracking paths with an effective score of not less than 0.60 are retained, and the event node with the earliest event time in the backtracking path is extracted as the candidate first event node. When there are multiple candidate first event nodes, the candidate first event node with the highest effective score is selected first. If the effective scores are the same, the candidate first event node with the earliest event time is selected. If the event time is still the same, the candidate first event node with the highest alarm credibility correction score is selected to determine the first event node.
[0165] By associating the initial event node, object node, and region node, a risk location result is generated that includes the risk object, the initial event node, the associated region node, and the location time period. Specifically, the association of the initial event node, object node, and region node is as follows:
[0166] Read the initial event node, corrected alarm records, risk transmission relationship table including the event time period, and security node table; using the initial event node as the association starting point, search the risk transmission relationship table including the event time period for risk transmission records where the starting or ending node is equal to the initial event node, and extract the object nodes and area nodes connected to the initial event node; read the object identifier and object category corresponding to the object node from the security node table, and write the object identifier as a risk object; read the area number and area sensitivity level corresponding to the area node from the security node table, and write the area number as the associated area node; calculate the location association score, multiply the risk transmission score by 0.45, and multiply the alarm credibility correction score in the corrected alarm record by 0.3. 5. Multiply the corresponding value of the regional sensitivity level by 0.20 and add them together to obtain the location association score. The corresponding values for regional sensitivity levels 1, 2, 3 and 4 are 0.25, 0.50, 0.75 and 1.00 respectively. Retain object nodes and regional nodes with a location association score of not less than 0.60. If there are multiple object nodes or regional nodes, they are arranged from high to low according to the location association score. The location time period is the union of the event time period of the first event node, the event time period of the correction alarm record and the event time period of the retained risk transmission record. The location time period is formed by the earliest event start point and the latest event end point. Combine the risk object, the first event node, the associated regional node, the location time period and the location association score to generate the risk location result.
[0167] In this embodiment, generating a security risk assessment result that includes the risk object, risk area, risk type, risk level, and handling priority includes:
[0168] The risk object, initial event node, associated regional nodes, and location time period are matched with regional sensitivity level, access permission constraints, deployment time period constraints, and risk handling level constraints to generate a basic risk assessment record. Specifically, the generation of this basic risk assessment record includes:
[0169] Read the risk object, initial event node, associated area node, and location time period from the risk location results. Also read the area sensitivity level, access permission constraints, deployment time period constraints, and risk handling level constraints associated with the area number from the security constraint data. Using the area number corresponding to the associated area node as the matching key, write the area sensitivity level; if no area sensitivity level is matched, write level 1 and mark it as missing. Match the risk object and area number against the list of authorized objects in the access permission constraints; if the risk object is included in the list, the access permission hit value is 0; otherwise, it is 1. Then, locate... The deployment hit value is set to 1 if the start time of the positioning time period is not later than the end time of the deployment and the end time of the positioning time period is not earlier than the start time of the deployment; otherwise, it is set to 0. The risk handling level constraint is matched with the alarm type and area number corresponding to the first event node, and the risk handling level is written. If no risk handling level is matched, the lowest handling level is written and marked as missing handling constraint. The risk object, the first event node, the associated area node, the positioning time period, the area sensitivity level, the access permission hit value, the deployment hit value, the risk handling level and the missing mark are combined to generate the basic record of risk assessment.
[0170] The event type corresponding to the initial event node is determined as the risk type, and the area corresponding to the associated area node is determined as the risk area. A risk assessment association record is then generated. Specifically, the generation of the risk assessment association record involves:
[0171] The system reads the risk object, initial event node, associated area node, and location time period from the basic risk assessment record, and reads the node attributes in the security node table that match the initial event node and associated area node. Using the node identifier of the initial event node as the matching key, it retrieves the event type corresponding to the event node, writes the event type as the risk type, and if the event type is empty, writes the alarm type as the risk type and marks it as type filling. Using the node identifier of the associated area node as the matching key, it retrieves the area location corresponding to the area node, writes the area location as the risk area, and if the area location is empty, writes the area number as the risk area. It performs deduplication on records where the risk object, risk type, risk area, and location time period are the same, and retains the record with the highest location association score. It combines the risk object, risk type, risk area, initial event node, associated area node, location time period, and type filling mark to generate a risk assessment association record.
[0172] Risk levels are determined based on risk type, regional sensitivity level, access control constraints, and deployment time constraints. Response priorities are then determined based on both the risk level and the risk response level constraints. Specifically, determining the risk level and the response priority based on these constraints is as follows:
[0173] Read the risk type, regional sensitivity level, access permission hit value, deployment hit value, and risk handling level from the basic risk assessment record and the associated risk assessment record. When the risk type and risk handling level constraints match successfully, convert the corresponding risk handling level into a basic risk type value: 0.25, 0.50, 0.75, and 1.00 for levels 1, 2, 3, and 4, respectively. If no match is found, the basic risk type value is 0.50. For regional sensitivity levels 1, 2, 3, and 4, the regional sensitivity values are 0.25, 0.50, 0.75, and 1.00, respectively. Multiply the basic risk type value by 0.40 and the regional sensitivity value by... The risk level score is obtained by adding 0.30, the access permission hit value multiplied by 0.20, and the deployment hit value multiplied by 0.10. A risk level score of not less than 0.80 is determined as high risk, not less than 0.60 and less than 0.80 is determined as medium-high risk, not less than 0.40 and less than 0.60 is determined as medium risk, and less than 0.40 is determined as low risk. The risk level and risk handling level are combined: high risk or level 4 handling level corresponds to level 1 priority, medium-high risk or level 3 handling level corresponds to level 2 priority, medium risk or level 2 handling level corresponds to level 3 priority, and low risk and level 1 handling level corresponds to level 4 priority.
[0174] Combine the risk object, risk area, risk type, risk level, handling priority, and location time period to generate a security risk assessment result.
[0175] refer to Figure 3 An intelligent security risk assessment system based on multi-source data fusion includes the following modules:
[0176] The data standardization module is used to process multi-source security data, historical false alarm data, and security constraint data to generate standard security data.
[0177] The trusted calibration module is used to perform false alarm calibration and cross-source consistency calibration on standard security data, and generate trusted security data and data source trustworthiness.
[0178] The event fusion module is used to merge target records, area records, and alarm records from trusted security data to generate trusted security event data.
[0179] The relationship graph construction module is used to construct a multi-source security relationship graph based on trusted security event data, data source credibility, and security constraint data.
[0180] The risk characterization module is used to input the multi-source security relationship diagram into the improved GATNE model and generate risk characterization vectors.
[0181] The risk transmission module is used to write back the risk representation vector to the object node, event node, and region node, and generate a risk transmission relationship table containing the event time period.
[0182] The risk location module is used to correct multi-source alarm conflicts based on the risk transmission relationship table and the reliability of the data source, and generate risk location results.
[0183] The assessment output module is used to generate security risk assessment results based on risk location results and security constraint data.
[0184] Example 1: To verify the feasibility of this invention in practice, it was applied to a security monitoring scenario in a smart park. The park includes office buildings, R&D buildings, underground parking garages, storage areas, and perimeter passages, with a total of 96 video cameras, 42 access control devices, 18 perimeter alarm devices, 36 fire detection devices, 64 environmental sensors, and 28 security device status acquisition terminals connected. The system collected 186,400 pieces of multi-source security data and 12,460 historical false alarm data over seven consecutive days. The security constraint data included 126 area codes, 386 access permission constraints, 54 deployment time constraints, and 32 risk handling level constraints.
[0185] After the data enters the processing flow, the system performs source identifier completion, field name unification, data format conversion, and anomaly sampling and removal on multi-source security data. Personnel ID, vehicle ID, area field, and alarm field are converted into unified object identifier, area ID, device ID, and event type. Before processing, approximately 11.8% of records had inconsistent time field formats, approximately 2.6% had missing device IDs, and approximately 6.4% had duplicate alarm records. After processing, 173,920 valid standard security data entries are retained. The system uses the security platform's time synchronization as a unified benchmark for time correction and generates time-series security data in fixed 30-second windows, reducing the average time deviation of cross-source records from 2.43 seconds to 0.46 seconds.
[0186] During the trusted calibration phase, the system generates a historical false alarm feature table based on historical false alarm data. Perimeter alarm device P011 recorded 86 false alarms between midnight and 6 AM, with 24 valid alarms, resulting in a historical false alarm suppression score of 0.218. Access control device D018 recorded 8 false alarms, with 96 valid alarms, resulting in a historical false alarm suppression score of 0.923. The system calculates the data source trustworthiness using weights of 0.35, 0.30, 0.20, and 0.15, removing 8421 low-trust records, ultimately generating 165499 trusted security data entries.
[0187] During the trusted security event fusion phase, taking the abnormal access to office building B2 as an example, video recognition records show that object O237 entered floor B2 at 8:32:14, access control records show that the same object failed to swipe its card at 8:32:16, and perimeter records show that the door zone abnormality was triggered at 8:32:20. The system calculates the target similarity as 1, the area overlap as 1, the alarm type consistency as 0.6, and the trust score as 0.87. After subtracting the time interval penalty, the event merging score is 0.842, which is greater than the merging threshold of 0.75, and therefore merged into the same trusted security event.
[0188] The system maps trusted security event data into a multi-source security relationship graph. Object records generate 18,420 object nodes, area records generate 126 area nodes, and alarm records generate 9,635 event nodes. Based on object-event relationships, event-area relationships, and object-area relationships, 58,360 trusted relationship edges are generated. Based on area sensitivity levels, access permission constraints, deployment time constraints, and device coverage constraints, 14,280 area constraint edges are generated.
[0189] In the improved GATNE model processing, the system inputs the multi-source security relationship graph into the security object attribute embedding layer to generate a security node embedding matrix with a size of 28181×128. After entering the trusted relationship edge embedding calibration layer, the system concatenates the starting node embedding vector, the ending node embedding vector, and the direction difference vector, and superimposes the edge trustworthiness and constraint labels to generate a trusted edge embedding matrix with a size of 72640×128.
[0190] In the relation type attention fusion layer, with event node E0832 as the central node, the object event edge score is 0.86, the event region edge score is 0.91, and the access permission constraint edge score is 0.94, corresponding to relation type weights of 0.317, 0.336, and 0.347, respectively. After weighted aggregation, the risk response value of event node E0832 increases from 0.62 to 0.79.
[0191] In the multi-granularity security risk projection layer, the system divides the relationship fusion node embedding matrix into object-level, region-level, and event-level granularities, generating 18,420 object risk embeddings, 126 region risk embeddings, and 9,635 event risk embeddings. After consistency calibration of connected nodes within the same event period, the average difference between the three types of embeddings within the same abnormal event group decreased from 0.182 to 0.071.
[0192] In the heterogeneous Skip-Gram output layer, the system uses object nodes, region nodes, and event nodes as the starting points for the walk, with a walk length of 6 nodes. The context window takes 2 nodes before and after the walk, generating context node prediction results, which are then aggregated to form a risk representation vector. During training, the context node prediction error decreased from 1.824 to 0.493, and the reliable relation edge reconstruction error decreased from 0.712 to 0.218.
[0193] The system writes the risk representation vector back to the object node, event node, and region node according to the node identifier. The node risk intensity value of object O237 is 0.86, the risk intensity value of event E0832 is 0.91, and the risk intensity value of region B2 is 0.78. After generating candidate risk transmission records along the trusted relationship edges and region constraint edges, the risk transmission score threshold is set to 0.65, and 21,480 records that meet the threshold are retained, forming a risk transmission relationship table containing the event time period.
[0194] During the alarm conflict correction phase, the system identified 326 sets of multi-source alarm conflict records. Taking object O237 as an example, the access control device issued an unauthorized access alarm, the video device issued an abnormal loitering alarm, and the perimeter device issued a door zone trigger alarm. The reliability correction score for the unauthorized access alarm was 0.89, the abnormal loitering alarm was 0.72, and the door zone trigger alarm was 0.68. Ultimately, the unauthorized access alarm was determined to be a valid alarm.
[0195] During the initial event tracing phase, the system backtracks along trusted relationship edges and regional constraint edges, with a tracing depth of 5 layers and a tracing time range of 10 30-second event periods. The risk corresponding to object O237 is traced back from the correction alarm record to the access control card swipe failure event E0829. The tracing validity score is 0.83, and E0829 is identified as the initial event node. The risk location results show that the risk object is O237, the associated regional node is B2, the location time is from 8:31:30 to 8:33:00, the risk type is unauthorized access risk, the risk level is high risk, and the handling priority is first-level priority.
[0196] In the comparative experiment, the system compared the method of this invention with the traditional fixed-weight fusion method. The test sample included 1200 sets of manually verified security events. The effective alarm identification accuracy rate was 81.6% for the traditional method and 93.4% for this invention; the alarm conflict correction accuracy rate was 74.2% for the traditional method and 90.8% for this invention; the first event tracing accuracy rate was 68.5% for the traditional method and 88.1% for this invention; and the high-risk event false negative rate was 12.7% for the traditional method and 4.3% for this invention.
[0197] As can be seen from Example 1, the present invention can complete the standardization of multi-source security data, calibration of data source credibility, fusion of credible security events, construction of multi-source security relationship diagram, generation of risk characterization and tracing of first-time events. It can effectively reduce the impact of false alarms and alarm conflicts on security risk assessment results, and verify the engineering feasibility of the present invention in smart park security scenarios.
[0198] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for intelligent security risk assessment based on multi-source data fusion, characterized in that, include: Collect multi-source security data, historical false alarm data, and security constraint data, and preprocess them to generate standard security data; Based on historical false alarm data and cross-source consistency calibration standard security data, generate trusted security data and data source credibility; Merge target records, area records, and alarm records from trusted security data to generate trusted security event data; A multi-source security relationship graph is constructed based on trusted security event data, data source credibility, and security constraint data. The multi-source security relationship graph includes object nodes, region nodes, event nodes, trusted relationship edges, and region constraint edges. The multi-source security relationship graph is input into the improved GATNE model to generate a risk representation vector. The improved GATNE model includes a security object attribute embedding layer, a trusted relationship edge embedding calibration layer, a relationship type attention fusion layer, a multi-granularity security risk projection layer, and a heterogeneous Skip-Gram output layer. Write the risk representation vector back to the object node, event node, and region node, and generate a risk transmission relationship table containing the event time period along the trusted relationship edge and the region constraint edge; Based on the risk transmission relationship table and the credibility of the data source, the conflict of multiple alarms from the same object, the same region, and the same time period is corrected, the first event node and related regional nodes are traced, and the risk location results are generated. Based on the risk location results and security constraint data, a security risk assessment result is generated, which includes the risk object, risk area, risk type, risk level, and handling priority.
2. The intelligent security risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The preprocessing generates standard security data including: The data items for multi-source security data, historical false alarm data, and security constraint data are identified, and source identifier completion, field name unification, data format conversion, and anomaly sampling and removal are performed. The multi-source security data includes video recognition data, access control data, perimeter alarm data, fire sensing data, environmental sensing data, and security equipment status data. The historical false alarm data includes false alarm event identifier, false alarm source device, false alarm occurrence area, false alarm occurrence time, false alarm trigger type, and false alarm handling result. The security constraint data includes area sensitivity level, access permission constraint, deployment time period constraint, equipment coverage range constraint, and risk handling level constraint. Based on the collection time, device timestamp, alarm trigger time, and handling time, the time benchmark is unified and the event time period is segmented to generate time-series security data; Based on object identifier, area number, device number, event type, and constraint type, associated time-series security data, historical false alarm data, and security constraint data, perform missing field filling, duplicate record merging, and sensitive field desensitization to generate standard security data.
3. The intelligent security risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The credibility of the generated trusted security data and the data source includes: Based on historical false alarm data, the number of false alarms, the number of valid alarms, and the number of times the handling results are consistent are counted according to the source device, event type, occurrence area, and occurrence time, and a historical false alarm feature table is generated. Perform event matching on video recognition records, access control records, perimeter alarm records, and environmental perception records for the same target, the same area, and the same time period in standard security data to generate cross-source consistent records; The data source credibility is obtained by weighted summation of historical false alarm suppression score, cross-source consistency score, complete collection score and time synchronization score. Based on the data source credibility, the standard security data is then subjected to credibility marking, conflict record downweighting and low credibility record removal to generate credible security data.
4. The intelligent security risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The generated trusted security event data includes: Candidate security event records are generated by combining target records, area records, and alarm records according to target identifier, area number, device number, alarm type, and event time period. The event merge score is obtained by subtracting the time interval penalty value from the weighted result of target similarity, region overlap, alarm type consistency and data source credibility. Candidate security event records that meet the merging threshold are grouped into the same trusted security event. The target record, area record, alarm record, source device, and event time period are merged to generate trusted security event data.
5. The intelligent security risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The construction of the multi-source security relationship diagram based on trusted security event data, data source credibility, and security constraint data includes: Map target records, area records, and alarm records to object nodes, area nodes, and event nodes, configure node types, node sources, and event time periods, and generate a security node table; Match the relationships between object nodes, region nodes, and event nodes to generate trusted relationship edges; Generate area constraint edges based on area sensitivity level, access permission constraints, deployment time constraints, and equipment coverage constraints; Encapsulate object nodes, region nodes, event nodes, trusted relationship edges, and region constraint edges to generate a multi-source security relationship graph.
6. The intelligent security risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The generated risk representation vector includes: Based on the GATNE model's basic embedding, edge embedding, global node embedding generation, and heterogeneous Skip-Gram output, an improved GATNE model is formed by adding a security object attribute embedding layer at the basic embedding input, a trusted relationship edge embedding calibration layer at the edge embedding branch, a relationship type attention fusion layer at the edge type self-attention combination position, and a multi-granularity security risk projection layer at the global node embedding generation output. The training objects of the heterogeneous Skip-Gram output layer are set as object nodes, region nodes, and event nodes. Input the multi-source security relationship diagram into the security object attribute embedding layer, and generate a security node embedding matrix according to node type, node source, event time period and node attribute; The security node embedding matrix, trusted relation edges, regional constraint edges, and edge credibility are input into the trusted relation edge embedding calibration layer, and a trusted edge embedding matrix is generated according to edge type, edge connection node, edge credibility, and constraint label. Input the security node embedding matrix and the trusted edge embedding matrix into the relation type attention fusion layer, and generate the relation fusion node embedding matrix according to the relation type of trusted relation edges and region constraint edges; The relationship fusion node embedding matrix is input into the multi-granularity security risk projection layer, and a multi-granularity risk embedding matrix is generated according to object granularity, region granularity and event granularity. The multi-granularity risk embedding matrix is input into the heterogeneous Skip-Gram output layer, and the context node prediction results are generated based on the node walk sequence. The multi-granularity risk embedding matrix is then aggregated to generate a risk representation vector. The improved GATNE model is trained by weighting the context node prediction error, the credible relation edge reconstruction error, and the multi-granularity risk embedding consistency error, so that the trained improved GATNE model can be used to generate risk representation vectors.
7. The intelligent security risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The process of generating a risk transmission relationship table containing event time periods along trusted relationship edges and regional constraint edges includes: The risk representation vector is written back to the object node, event node, and region node according to the node identifier, generating a node risk write-back table; Traverse trusted relation edges and region constraint edges, combine edge connection direction, edge type, edge trustworthiness and constraint label to generate candidate risk transmission records; Calculate the risk transmission score of the candidate risk transmission record. The risk transmission score is obtained by subtracting the event time interval penalty value from the weighted average of the risk representation similarity, edge credibility and constraint matching value between the starting node and the ending node. Candidate risk transmission records that meet the transmission threshold are selected and organized into a risk transmission relationship table containing the event time period.
8. The intelligent security risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The generated risk location results include: Multi-source alarm records in the risk transmission relationship table are aggregated by object node, region node, and event time period, and alarm conflict records within the same object node, the same region node, and the same event time period are marked. By combining the data source credibility and risk transmission relationship table, a reliable correction is performed on the alarm conflict records to generate corrected alarm records; Backtrack along trusted relation edges and region constraint edges to correct the event nodes corresponding to alarm records and determine the first event node; Associate the initial event node, object node, and region node to generate a risk location result that includes the risk object, the initial event node, the associated region node, and the location time period.
9. The intelligent security risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The security risk assessment results generated include risk objects, risk areas, risk types, risk levels, and handling priorities. Match the risk object, the initial event node, the associated regional nodes and the location time period with the regional sensitivity level, access permission constraints, deployment time constraints and risk handling level constraints to generate a basic record for risk assessment; The event type corresponding to the first event node is determined as the risk type, and the area location corresponding to the associated area node is determined as the risk area, and a risk assessment association record is generated. Risk levels are determined by risk type, regional sensitivity level, access control constraints, and deployment time constraints; and disposal priorities are determined by risk level and risk disposal level constraints. Combine the risk object, risk area, risk type, risk level, handling priority, and location time period to generate a security risk assessment result.
10. An intelligent security risk assessment system based on multi-source data fusion, comprising executing the intelligent security risk assessment method based on multi-source data fusion as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The data standardization module is used to process multi-source security data, historical false alarm data, and security constraint data to generate standard security data. The trusted calibration module is used to perform false alarm calibration and cross-source consistency calibration on standard security data, and generate trusted security data and data source trustworthiness. The event fusion module is used to merge target records, area records, and alarm records from trusted security data to generate trusted security event data. The relationship graph construction module is used to construct a multi-source security relationship graph based on trusted security event data, data source credibility, and security constraint data. The risk characterization module is used to input the multi-source security relationship diagram into the improved GATNE model and generate risk characterization vectors. The risk transmission module is used to write back the risk representation vector to the object node, event node, and region node, and generate a risk transmission relationship table containing the event time period. The risk location module is used to correct multi-source alarm conflicts based on the risk transmission relationship table and the reliability of the data source, and generate risk location results. The assessment output module is used to generate security risk assessment results based on risk location results and security constraint data.
