A dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs

By employing a dual-channel threat retrieval and multi-agent collaborative analysis method, the problems of incomplete coverage and unverifiable analysis conclusions in a single retrieval channel within a computing power network hub are solved. This method enables the joint extraction of explicit threat entities and implicit attack relationships, providing verifiable threat analysis conclusions.

CN123123031APending Publication Date: 2026-11-17GUIZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611613496.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing computing network hubs, a single retrieval channel cannot simultaneously retain explicit threat entities and implicit attack relationships, resulting in incomplete coverage of context learning examples. Selecting examples based on relevance is prone to semantic redundancy, consuming limited model input space. There is a lack of stable structured connections between unstructured threat intelligence and standardized attack knowledge, leading to insufficient verifiability and interpretability of analysis conclusions.

Method used

A dual-channel threat retrieval method is adopted, combining entity-oriented sparse retrieval and relation-oriented dense retrieval. The maximum marginal relevance selection score suppresses example redundancy, constructs a current threat knowledge subgraph, and conducts collaborative analysis in a shared inference storage area. Multi-agent coordination is used to verify evidence consistency and interpret the output of attack chains, attack technology mappings, and attack organization profiles.

Benefits of technology

It expands the coverage of entity and relationship types within a limited context space, reduces interference from duplicate cases, provides verifiable and interpretable threat analysis conclusions, and ensures the traceability of attack chains, attack technique mappings, and attack organization profiles.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123031A_ABST
    Figure CN123123031A_ABST
Patent Text Reader

Abstract

This invention discloses a dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs, relating to the field of network security technology. It includes: acquiring threat intelligence to be analyzed; obtaining a candidate pool through sparse and dense retrieval channels; selecting a set of context examples by fusing dual-channel relevance and combining the maximum marginal relevance selection score; constructing a current threat knowledge subgraph by jointly extracting entities and semantic relationships under the constraints of a preset threat intelligence ontology execution mode; obtaining a recall knowledge set from heterogeneous security knowledge sources; having a coordinating agent invoke various specialized agents; verifying the consistency of evidence for intermediate conclusions executed by the adjudicating agents and resolving conflicts; and finally, having an interpreting agent output evidence-supported attack chains, attack technique mappings, attack organization profiles, and a threat analysis report. This invention takes into account both explicit threat entities and implicit attack relationships, reduces contextual redundancy, and provides verifiable evidence for the analysis conclusions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the technical fields of network security, knowledge graphs, and multi-agent systems, and in particular, to a dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs. Background Technology

[0002] As computing network hubs continuously expand the exposure of data centers, cloud computing platforms, edge connections, and supply chain components, attackers can leverage exposed services, credentials, scheduling interfaces, or inter-node links to conduct persistent penetration. Security incidents can easily spread along computing tasks, data links, and management relationships. The network summaries, operation logs, alerts, and vulnerability reports aggregated by hubs are often heterogeneous and unstructured. Manually extracting attack organizations, tools, malware, vulnerabilities, targets, and their behavioral relationships from these data is costly and fails to meet the needs of continuous analysis and rapid correlation.

[0003] When using contextual learning for threat intelligence extraction, existing methods typically retrieve examples only within a single representation space. Sparse retrieval based on keywords or term frequency can expose explicit entity clues such as attack organization names, malware names, tool names, and vulnerability numbers, but it struggles to uncover historical cases with similar attack relationships despite different expressions. Dense retrieval based on semantic vectors, while capturing implicit behavioral and relational similarities, may overlook entity terms with discriminative power. Selecting examples solely based on relevance introduces highly repetitive examples into the context, crowding out the limited model input space and interfering with the joint extraction of threat entities and semantic relationships.

[0004] Furthermore, entity or relation triples extracted from unstructured threat intelligence typically lack stable, structured connections with standardized attack knowledge bases, historical attack cases, and original evidence. Some solutions directly output attack attribution, attack technique mapping, or question-and-answer conclusions after obtaining the triples, without jointly verifying the threat knowledge graph, external security knowledge, and evidence. This makes it difficult to uniformly generate attack chain reconstruction, attack technique mapping, and attack organization profiling, resulting in insufficient traceability and interpretability of the analysis conclusions. When tasks such as attack chain reconstruction, attack mapping, attack organization profiling, evidence verification, and interpretation generation are performed separately by different models or agents, the lack of shared reasoning states and unified adjudication mechanisms may lead to conflicting intermediate conclusions, or even attribution results lacking supporting evidence.

[0005] In summary, to address the shortcomings of existing technologies, such as the inability of a single retrieval channel to simultaneously retain explicit threat entities and implicit attack relationships, resulting in incomplete example coverage for context learning; the tendency to directly select examples based on relevance easily leads to semantic redundancy, consuming limited model input space and interfering with joint entity relationship extraction; the lack of stable structured connections between unstructured threat intelligence and standardized attack knowledge, making it difficult to uniformly generate attack chains, attack technique mappings, and attack organization profiles; and the lack of shared reasoning states, heterogeneous evidence verification, and conflict resolution among multiple analysis units, resulting in insufficient verifiability and interpretability of threat analysis conclusions, a dual-channel threat retrieval and multi-agent collaborative analysis method oriented towards computing power hubs is proposed. Summary of the Invention

[0006] The main objective of this invention is to provide a dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs, which takes into account both explicit threat entities and implicit attack relationships, suppresses contextual redundancy, and outputs verifiable and interpretable threat analysis conclusions such as attack chains, attack technology mappings, and attack organization profiles through pattern-constrained entity relationship joint extraction, knowledge retrieval, shared reasoning, and evidence adjudication.

[0007] Based on the first main aspect of the present invention, a dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs is provided, comprising the following steps:

[0008] Acquire threat intelligence and asset context to be analyzed from the computing power network hub, identify the threat intelligence to be analyzed as query text, and obtain a unified example library consisting of annotated threat intelligence examples;

[0009] The corresponding candidate example sets are obtained from the unified example library through the entity-oriented sparse retrieval channel and the relation-oriented dense retrieval channel respectively, and the union of these sets is used to obtain the candidate pool.

[0010] By integrating dual-channel relevance and combining the maximum marginal relevance to select scores and suppress example redundancy, a set of context examples is selected from the candidate pool, and a current threat knowledge subgraph is constructed based on the joint extraction of entities and semantic relationships under the constraints of the context example set and the preset threat intelligence ontology execution mode.

[0011] Based on the current threat knowledge subgraph, a recall knowledge set is obtained from heterogeneous security knowledge sources, and the recall knowledge set and the current threat knowledge subgraph are written into the shared inference storage area.

[0012] The coordinating agent invokes each dedicated agent based on the current threat knowledge subgraph, the recalled knowledge set, and the preset task routing rules; the adjudicating agent performs evidence consistency verification and conflict resolution of the intermediate conclusions of each dedicated agent based on the shared inference storage area.

[0013] Based on the evidence consistency verification results, the interpreting agent outputs the attack chain, attack technology mapping, attack organization profile, and threat analysis report supported by the evidence.

[0014] From an overall conceptual standpoint, this invention utilizes two complementary perspectives—entity terms and relational semantics—to form candidate contexts during the extraction phase, and retains examples with information gain through redundancy penalties. During the analysis phase, the current threat knowledge subgraph, external security knowledge, and multi-agent reasoning results are written into a shared reasoning storage area, where an adjudicating agent performs evidence gating. This connects context retrieval, structured extraction, and evidence analysis into a unified processing flow. Through this technical solution, the previously fragmented retrieval, extraction, knowledge retrieval, and multi-agent analysis are linked into a closed loop, solving problems such as incomplete coverage of single retrieval channels, example redundancy, difficulty in structuring unstructured intelligence, lack of shared reasoning and conflict resolution among multiple analysis units, and untraceable conclusions. This achieves complementary retrieval contexts, consistent extraction patterns, and verifiable collaborative analysis conclusions.

[0015] As a further preferred embodiment, in the aforementioned method, the entity-oriented sparse retrieval channel calculates sparse relevance using a term frequency-inverse document frequency sorting method:

[0016] ;

[0017] in, Indicates sparse correlation; Indicates the query text; Indicates candidate examples; Indicates terms; Indicates the inverse document frequency of a term; Indicates the frequency of a term in the candidate examples; This represents the preset parameters for controlling word frequency saturation. This represents the preset parameters for controlling length normalization. This represents the average length of examples in the Uniform Example Library;

[0018] Select the top results based on sparsity correlation, from highest to lowest. From 10 candidate examples, a sparse candidate set is obtained;

[0019] The relation-oriented dense retrieval channel uses a security domain text encoding model to generate query vectors and candidate vectors, and calculates dense relevance based on cosine similarity.

[0020] Among them, the text encoding model in the security domain adopts an encoding model pre-trained or adapted with a cybersecurity corpus; and the top models are selected according to the density of relevance from high to low. 10 candidate examples are used to obtain a dense candidate set.

[0021] In the above technical solution, two complementary retrieval channels are used to achieve coverage of the perspective of entities and relationships. The sparse channel retains explicit entity clues such as the name of the attacking organization, the name of the malware, the name of the tool, and the vulnerability number, while the dense channel captures the attack behavior and relational semantics. Thus, in the same query, entity matching examples and relational similarity examples can be obtained at the same time, avoiding incomplete context coverage caused by a single retrieval perspective.

[0022] As a further preferred embodiment, in the aforementioned method, the execution steps for selecting the context example set from the candidate pool are as follows:

[0023] Normalize the sparse relevance and dense relevance in the candidate pool; for candidate examples and selected context example sets, calculate the redundancy of the candidate example relative to the selected context example set. When the selected context example set is empty, the redundancy is zero; when the selected context example set is not empty, the redundancy is the maximum value of the secure semantic similarity between the candidate example and any selected example in the selected context example set.

[0024] The maximum marginal relevance selection score is calculated by combining redundancy:

[0025] ;

[0026] in, This represents the score for the selection of the maximum marginal relevance. Indicates the query text; Indicates candidate examples; This indicates the selected context example collection; Weighting coefficients representing the degree of sparse correlation; Weighting coefficients representing the degree of close correlation; Indicates the redundancy penalty weight; Indicates redundancy.

[0027] In each round, candidate examples whose scores meet the preset conditions are added to the selected context example set, until a preset number of context examples are obtained or until the candidate pool is empty.

[0028] In selecting the context example set, the dual-channel relevance and the redundancy between examples are considered in a unified manner. This avoids the situation where selecting examples based solely on relevance leads to highly repetitive content that crowds out the limited model input space. As a result, the coverage of entity types, relationship types, and behavior patterns is expanded within the limited context space window, reducing the interference of duplicate examples on the joint extraction of entity relationships.

[0029] As a further preferred embodiment, in the aforementioned method, the steps for constructing the current threat knowledge subgraph are as follows:

[0030] The task instructions, preset threat intelligence ontology, structured output constraints, context example set and query text are combined into context learning prompts based on a preset form. The prompts are then input into a language model adapted to the security domain. Threat entities and semantic relationships are jointly extracted under the pattern constraints of the preset threat intelligence ontology. The extraction results are then subjected to type checking, alias alignment, duplicate item merging and confidence calculation, and finally assembled into the current threat knowledge subgraph.

[0031] The language model adapted for the security domain can learn threat entity boundaries, relational semantics, and triple patterns by adopting a low-rank adaptation method while keeping the pre-trained parameters frozen.

[0032] The entity types defined by the preset threat intelligence ontology include at least attackers, tools, malware, vulnerabilities, targets, and regions, and the relationship types include at least use, instruction, exploitation, attack target, and location.

[0033] The structured output constraint limits the output to machine-readable data containing triples of entity, entity type, and semantic relation.

[0034] Through the above technical solutions, unstructured threat text is converted into a consistent graph structure before entering multi-agent reasoning, reducing the impact of entity type confusion, relational ambiguity, and illegal output on subsequent analysis. This solves the problem of the lack of stable structured connections between structured threat intelligence and standardized attack knowledge, and provides a unified structural foundation for subsequent attack chain reconstruction, attack mapping, and attack organization profiling.

[0035] As a further preferred option, in the aforementioned method, the execution steps for obtaining the recall knowledge set from heterogeneous security knowledge sources are as follows:

[0036] The heterogeneous security knowledge sources on which the recall knowledge set is based include at least one of the following: threat intelligence knowledge graph, standardized attack tactics and techniques knowledge base, historical attack case library, evidence library, analysis rule library, computing power node asset ledger, node connection relationship and scheduling control plane log;

[0037] The knowledge recall controller uses entities, semantic relationships, and adjacency structures in the current threat knowledge subgraph as query conditions;

[0038] Calculate the correlation degree between candidate knowledge items and the current threat knowledge subgraph, sort them from high to low, select the top preset number of candidate knowledge items that meet the preset correlation degree conditions from heterogeneous security knowledge sources, and record the source, time, identifier and correlation score of the candidate knowledge items to obtain the recall knowledge set.

[0039] In the above technical solution, the current threat knowledge subgraph is used as an index to link standardized attack knowledge, historical cases and evidence in external heterogeneous security knowledge, providing traceable external evidence for multi-agent analysis and solving the problem of lack of stable connection between unstructured intelligence and standardized attack knowledge and lack of evidence support for subsequent analysis.

[0040] As a further preferred embodiment, in the aforementioned method, the coordinating agent reads the shared inference storage area and calls the attack organization profiling agent, attack technology mapping agent, evidence agent, inference agent, and explanation agent according to the task objectives and dependencies.

[0041] Each specialized intelligent agent can perform graph query, knowledge retrieval, relation expansion, rule verification, and text generation through a pre-set skill library and tool interface;

[0042] The adjudicating agent maintains facts, evidence, subgraphs, and intermediate results in the shared inference storage area. For inconsistent conclusions output by different agents, the adjudicating agent compares the consistency between the conclusion and the evidence in the current threat knowledge subgraph, the recall knowledge set, and the preset analysis rules, and performs retention, correction, rerouting, or rejection according to the sufficiency of the evidence.

[0043] In the above technical solution, complex threat analysis is broken down into collaborative tasks with clearly defined roles. By sharing reasoning states and using a unified adjudication mechanism, contradictory results generated independently by each specialized agent are avoided. This enables attack chains, attack technique mappings, attack organization profiles, and explanatory texts to be traced back to evidence items related to corresponding entity relationships. As a result, the verifiability and interpretability of threat analysis conclusions are not adequately addressed due to the lack of shared reasoning states, heterogeneous evidence verification, and conflict resolution among multiple analysis units.

[0044] As a further preferred embodiment, in the aforementioned method, when performing evidence consistency verification, the adjudicating agent only writes the candidate attack chain node, attack technique mapping, or attack organization attribute into the evidence consistency verification result if it is supported by facts in the current threat knowledge subgraph, evidence in the recall knowledge set, or both.

[0045] The method also includes writing the intermediate conclusions of each dedicated agent that have been manually reviewed back to the historical case library or analysis rule library, and recording the dedicated agent outputs that have been accepted, corrected or rejected.

[0046] The content generated by each dedicated intelligent agent is not directly written into the executable disposal strategy without being verified for evidence consistency by the adjudicating intelligent agent or manually reviewed.

[0047] In the above technical solution, by setting clear evidence gating conditions, candidate conclusions that are not supported by evidence in the current real-time threat facts or the knowledge recall set are excluded from the final verification results. This ensures that the attack chain, attack technology mapping, and organization profile that enter the final conclusion output can be traced back to the corresponding facts or evidence, preventing unfounded attack attribution and conclusion drift. Furthermore, the manual review forms a feedback and update loop, providing a basis for subsequent adjustments to task routing, evidence thresholds, and knowledge recall ranking. At the same time, the double review threshold before the disposal strategy is written avoids the generation of unverified or unreviewed content from directly affecting the disposal strategy, reducing the risk of mishandling.

[0048] Based on a second key aspect of the present invention, a dual-channel threat retrieval and multi-agent collaborative analysis system for implementing the aforementioned method, oriented towards a computing power network hub, is provided, comprising:

[0049] The data access module is used to acquire threat intelligence and asset context to be analyzed from the computing power network hub;

[0050] A dual-channel retrieval module is used to generate a candidate pool through an entity-oriented sparse retrieval channel and a relationship-oriented dense retrieval channel;

[0051] The example selection module is used to select a set of context examples based on the dual-channel relevance combined with the maximum marginal relevance score.

[0052] The joint extraction module is used to construct contextual learning prompts and jointly extract threat entities and semantic relations under pattern constraints;

[0053] The knowledge graph module is used to normalize the extracted results and construct a knowledge subgraph of the current threat.

[0054] The knowledge retrieval module is used to retrieve a knowledge set from heterogeneous security knowledge sources and obtain the retrieved knowledge set.

[0055] A multi-agent collaboration module is used to allocate analysis tasks to multiple specialized agents by coordinating agents;

[0056] The adjudication and shared storage module is used to maintain the shared inference storage area and perform evidence consistency verification and conflict resolution;

[0057] The analysis output module is used to output evidence-supported attack chains, attack technique mappings, attack organization profiles, and threat analysis reports.

[0058] Based on a third key aspect of the present invention, an electronic device is provided, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0059] The memory stores a computer program that, when executed by the processor, causes the processor to perform the dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs as described above.

[0060] Based on the fourth principal aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed, implements the aforementioned dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs.

[0061] Compared with existing technologies, this invention provides a dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs. First, this invention achieves dual-channel complementarity and context optimization at the retrieval and extraction levels. Entity-oriented sparse retrieval retains explicit entity clues such as attack organization names, malware names, tool names, and vulnerability numbers, while relationship-oriented dense retrieval discovers historical cases with different expressions but similar attack behaviors or relational semantics, thus avoiding incomplete example coverage caused by a single retrieval channel. Simultaneously, the fusion of dual-channel relevance introduces maximum marginal relevance to suppress example redundancy, reducing duplicate cases within a limited context space and expanding the coverage of entity types, relationship types, and attack patterns. This provides a complementary context for subsequent joint entity-relation extraction. Furthermore, under the constraints of a preset threat intelligence ontology and structured output, joint entity-relation extraction is performed, and after normalization processing, unstructured threat intelligence is stably converted into a pattern-consistent current threat knowledge subgraph, providing a unified graph structure for subsequent processing.

[0062] Secondly, at the analysis level, this invention writes the current threat knowledge subgraph and knowledge recalled from heterogeneous security knowledge sources such as standardized attack knowledge bases, historical attack case databases, and evidence databases into a shared inference storage area. A coordinating agent distributes attack chain reconstruction, attack technique mapping, attack organization profiling, evidence verification, and interpretation generation tasks to multiple specialized agents. An adjudicating agent then performs evidence consistency verification and conflict resolution based on the shared inference storage area. Thus, attack chain nodes, attack technique maps, and attack organization profiling can only reach the final conclusion after obtaining support from current threat facts or external evidence. This allows attack chains, attack technique maps, attack organization profiling, and threat analysis reports to be traced back to corresponding entity relationships, knowledge items, and intermediate structures, improving the verifiability and interpretability of the analysis conclusions and avoiding conclusion drift and contradictions in the reasoning processes of multiple analysis units.

[0063] Finally, overall, this invention connects dual-channel threat retrieval, contextual example selection, pattern-constrained entity relationship joint extraction, threat knowledge graph construction, heterogeneous knowledge retrieval, and multi-agent collaborative analysis into a unified processing flow, taking into account both explicit threat entities and implicit attack relationships, reducing contextual redundancy, and ensuring that the analysis conclusions have verifiable evidence. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0065] Figure 1 The following is an execution flowchart of a dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs, according to one embodiment of the present invention.

[0066] Figure 2 The diagram illustrates the overall architecture of dual-channel context retrieval, example selection, threat intelligence extraction, and multi-agent analysis in one embodiment of the present invention.

[0067] Figure 3 This diagram illustrates an example flowchart of dual-channel threat retrieval and maximum marginal relevance selection in one embodiment of the present invention.

[0068] Figure 4 This diagram illustrates a flowchart of guided entity relationship extraction and threat knowledge graph construction in one embodiment of the present invention.

[0069] Figure 5 A flowchart illustrating knowledge recall enhancement multi-agent collaborative analysis and evidence output is shown in one embodiment of the present invention. Detailed Implementation

[0070] The preferred embodiments of the present invention will be described in detail below to provide a clearer understanding of the purpose, features, and advantages of the invention. It should be understood that the following embodiments are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the technical solution of the invention.

[0071] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known techniques associated with the invention may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0072] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0073] The specific meanings of the technical terms or English abbreviations that may be used in this invention are explained as follows:

[0074] Computing network hub: A hub node or hub platform used to aggregate information from data centers, cloud computing platforms, edge nodes, network connections and scheduling control, and to orchestrate or collaboratively manage computing tasks across nodes.

[0075] Threat intelligence to be analyzed: natural language text and its metadata from security reports, threat intelligence subscriptions, alert logs, event logs, vulnerability announcements, malicious sample analysis results, or manual assessment records.

[0076] Asset context: Information such as computing centers, computing nodes, scheduling platforms, management interfaces, business systems, network connections, regions, business importance, and security policies associated with the threat intelligence to be analyzed.

[0077] Unified Example Library: A collection of retrievals consisting of labeled threat intelligence examples. Each example contains at least the original intelligence text, and one of the following: the corresponding entity, entity type, semantic relation, or relation triplet.

[0078] Entity-oriented sparse retrieval: Calculate the relevance between query text and examples based on the frequency and distinctiveness of terms such as attack organization name, malware name, tool name, vulnerability number, and target name.

[0079] Relationship-oriented dense retrieval: Using a security-domain text encoding model, query text and examples are encoded into vectors, and examples with similar attack behaviors or semantic relationships are discovered based on vector similarity.

[0080] Contextual Examples: Annotated threat intelligence examples selected from the candidate pool are used to provide entity boundaries, relationship patterns, and structured output demonstrations to the joint extraction model in contextual learning cues.

[0081] Threat intelligence ontology: A predefined set of patterns for the types of threat entities and semantic relationships that can be extracted. Entity types may include attackers, tools, malware, vulnerabilities, targets, regions, organizations, and industries; relationship types may include use, development, attribution, instruction, exploitation, attack target, location, and impact.

[0082] Current Threat Knowledge Subgraph: A graph structure consisting of normalized entities as nodes and normalized semantic relationships as edges in the current threat intelligence to be analyzed.

[0083] Shared reasoning storage area: A structured storage space for recording facts, evidence, subgraphs, task states, intermediate conclusions, conflict items, and adjudication results during multi-agent collaborative analysis.

[0084] Evidence gating: A control mechanism that allows candidate conclusions to enter the final analysis results only when supported by the current threat knowledge subgraph, the recalled knowledge set, or preset analysis rules.

[0085] Combination Figure 1 As shown, this invention provides a dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs, comprising the following steps:

[0086] S100, obtain the threat intelligence and asset context to be analyzed gathered by the computing power network hub, determine the threat intelligence to be analyzed as query text, and obtain a unified example library consisting of annotated threat intelligence examples;

[0087] S200, respectively obtain the corresponding candidate example sets from the unified example library through the entity-oriented sparse retrieval channel and the relation-oriented dense retrieval channel, and take the union of them to obtain the candidate pool;

[0088] S300 integrates dual-channel relevance and combines the maximum marginal relevance to select scores and suppress example redundancy. It selects a set of context examples from the candidate pool and constructs a current threat knowledge subgraph based on the joint extraction of entities and semantic relationships under the constraints of the context example set and the preset threat intelligence ontology execution mode.

[0089] S400, Based on the current threat knowledge subgraph, obtain the recall knowledge set from the heterogeneous security knowledge source, and write the recall knowledge set and the current threat knowledge subgraph into the shared inference storage area;

[0090] S500: The coordinating agent calls each dedicated agent based on the current threat knowledge subgraph, the recalled knowledge set, and the preset task routing rules; the adjudicating agent performs evidence consistency verification and conflict resolution of the intermediate conclusions of each dedicated agent based on the shared inference storage area.

[0091] S600, based on the evidence consistency verification results, outputs an attack chain, attack technique mapping, attack organization profile, and threat analysis report supported by evidence from the interpreting agent.

[0092] Combination Figure 2As shown, the overall system for implementing the method of this invention includes four stages in the processing order: dual-channel context retrieval, example selection, prompt construction and threat intelligence extraction, and agent-based attack technique reasoning and interpretation.

[0093] The first phase involves parallel execution of entity-oriented sparse retrieval and relation-oriented dense retrieval.

[0094] The second stage utilizes the maximum marginal relevance score to select high-quality contextual examples.

[0095] The third stage generates entities, relations, and triples using a joint extraction model.

[0096] In the fourth stage, based on the threat knowledge graph, standardized attack knowledge, historical attack cases, and evidence database, multiple intelligent agents collaboratively output interpretable attack organization profiles and attack technology analysis results.

[0097] In a feasible implementation scheme, step S100 is described in detail as follows:

[0098] The system receives a threat intelligence message describing an attacking organization, tool, malware, vulnerability, or target object, and combines the title, summary, body fragment, and asset context into a query text, attacking organization alias, and tool name in the order of preset fields, preserving the original character format.

[0099] The unified example library is a retrieval collection consisting of labeled threat intelligence examples. Each example contains at least the original intelligence text, as well as one of the following: the corresponding entity, entity type, semantic relationship, or relationship triple.

[0100] The main purpose of this step is to establish query inputs corresponding to the current security events in the computing power network hub.

[0101] Receive threat reports, alarm summaries, event logs, or analysis issues from the security data bus, log platform, intelligence platform, or analyst terminal, and read asset context such as computing nodes, management interfaces, connection paths, and business importance involved in the event.

[0102] The system can perform character encoding unification, invisible character filtering, and paragraph segmentation on text, but retain security semantics such as attack organization name, tool name, malware name, vulnerability number, command fragment, and region name.

[0103] In a feasible implementation scheme, step S200 is described in detail as follows:

[0104] The purpose of this step is to construct a candidate pool from two complementary perspectives. Figure 3As shown, the sparse retrieval channel performs exact matching for explicit entity terms, while the dense retrieval channel performs vector matching for attack behaviors and relational semantics. The two channels can be executed in parallel, each returning corresponding candidate examples; after deduplication based on the example identifiers, the union of these examples forms the candidate pool.

[0105] The sparse retrieval channel uses BM25 or other term ranking algorithms to obtain entity-related examples, while the dense retrieval channel uses SecureBERT or other security domain coding models pre-trained or adapted with cybersecurity corpora to obtain relation-related examples.

[0106] Specifically, for entity-oriented sparse retrieval, this invention uses a term frequency-inverse document frequency sorting method to calculate sparse relevance:

[0107] ;

[0108] in, Indicates sparse correlation; Indicates the query text; Indicates candidate examples; Indicates terms; Indicates the inverse document frequency of a term; Indicates the frequency of a term in the candidate examples; This represents the preset parameters for controlling word frequency saturation. This represents the preset parameters for controlling length normalization. This represents the average length of examples in the unified example library.

[0109] Select the top results based on sparsity correlation, from highest to lowest. 10 candidate examples are used to obtain a sparse candidate set.

[0110] For relation-oriented dense retrieval, this invention uses a security domain text encoding model to obtain query vectors and example vectors respectively, and returns candidate examples describing similar attack behaviors, delivery relationships, exploitation relationships or target relationships through cosine similarity. Even if the specific entities in two texts are different, as long as their attack relationships are similar, they can enter the dense candidate set.

[0111] Cosine similarity formula:

[0112] ;

[0113] in, Indicates the degree of dense correlation; Indicates the query text; Indicates candidate examples; Represents the query vector; This represents a candidate vector.

[0114] Selected from high to low density correlation 10 candidate examples are used to obtain a dense candidate set.

[0115] After obtaining the sparse candidate set and the dense candidate set respectively, the union of the two sets is used to form the candidate pool. During merging, the examples returned by the two channels can be deduplicated by unique identifier, while retaining the sparse relevance, dense relevance, and source channel information of each example.

[0116] If an example is hit by both channels, its scores for both channels are retained.

[0117] In one feasible embodiment, step S300 is described in detail as follows:

[0118] First, the sparse and dense relevance in the candidate pool are normalized. Then, for both the candidate example and the selected context example set, the redundancy of the candidate example relative to the selected context example set is calculated. .

[0119] When the selected context example set is empty, the redundancy is zero; when the selected context example set is not empty, the redundancy is the maximum value of the secure semantic similarity between the candidate example and any selected example in the selected context set.

[0120] ;

[0121] in, Indicates redundancy; This indicates the selected context example collection; Indicates candidate examples; Indicates that an example has been selected; Indicates similarity; This represents the maximum value.

[0122] Then, the maximum marginal relevance selection score of the candidate examples is calculated according to the following formula:

[0123] ;

[0124] in, This represents the score for the selection of the maximum marginal relevance. Indicates the query text; Indicates candidate examples; This indicates the selected context example collection; Weighting coefficients representing the degree of sparse correlation; Weighting coefficients representing the degree of close correlation; Indicates the redundancy penalty weight; Indicates redundancy.

[0125] In each round, the highest-scoring candidate example is selected from the remaining candidate examples and added to the set of selected context examples. And update the redundancy until the selected context example set contains The context example or candidate pool is empty.

[0126] The main purpose of the above process is to reduce redundancy between contextual examples. In the first round, examples are selected based on dual-channel fusion relevance; in subsequent rounds, the maximum similarity between each unselected candidate example and the selected set is calculated as a redundancy penalty, and the example with the highest maximum marginal relevance score is selected.

[0127] For example, when a query text simultaneously contains the name of an attack organization, a script tool, DNS data collection behavior, malware delivery by a loader, and a vulnerability number, a sparse channel can retrieve examples containing the same attack organization, tool, or vulnerability number, while a dense channel can retrieve similar examples describing relationships such as "malware delivery by a loader" or "access to a target after exploitation." After redundancy penalty, the system retains different entity clues and different relationship patterns, rather than repeatedly selecting multiple examples that only describe the same attack organization.

[0128] Through this embodiment, the system can simultaneously provide entity-level term correspondence and relation-level semantic correspondence in a limited context space, and reduce duplicate cases, providing complementary context for subsequent joint extraction of entity relations.

[0129] This embodiment also provides a joint extraction method that combines contextual examples with threat intelligence pattern constraints. Figure 4 As shown, the inputs to this method include query text, a set of context examples, task instructions, threat intelligence ontology, and structured output constraints.

[0130] The threat intelligence ontology restricts the types of entities and relationships that can be output, and the structured output constraints restrict the machine-readable data structure.

[0131] First, define entity types and relationship types. For example, entity types include attacker, tool, malware, vulnerability, target, and region; relationship types include use, instruction, exploitation, attack target, and location. Depending on the actual needs of the computing power network hub, entity types such as computing power nodes, scheduling platform, management interface, and connection links can also be added.

[0132] The machine-readable data structure consists of a triplet containing an entity, an entity type, and a relation.

[0133] The prompt template is populated in the order of task instructions, threat intelligence ontology, structured output constraints, context example set, and query text.

[0134] ;

[0135] in, This indicates a contextual learning prompt; Indicates the query text; Indicates task instructions; This represents the pre-defined threat intelligence entity; Indicates structured output constraints; This indicates a collection of context examples; it also indicates that the data is structured and assembled according to a preset template.

[0136] Each context example provides both the input text and the correct structured output, enabling the joint extraction model to reference entity boundaries and relationship patterns relevant to the current query.

[0137] Then, joint extraction of security domains is performed. In this embodiment, the LLaMA3-8B model with low-rank adaptation is used. Alternatively, a language model that can perform joint extraction of entity relations under preset ontology and output mode constraints can also be used.

[0138] This security-domain-adapted language model can learn threat entity boundaries, relational semantics, and triple patterns by using a low-rank adaptation method while keeping pre-trained parameters frozen.

[0139] The parameter update format is as follows:

[0140] ;

[0141] in, This represents the updated parameter matrix; Represents the pre-trained parameter matrix; Indicates the parameter update increment; and It is a trainable low-rank matrix.

[0142] By keeping the pre-trained model parameters frozen and training only the low-rank update matrix, the model learns the output methods of security domain terminology, entity boundaries, and relation triples. Under the pattern constraints of a pre-defined threat intelligence ontology, the model jointly extracts threat entities and semantic relationships.

[0143] During reasoning, the pattern constraint decoding unit restricts entity types and relation types to come from a predefined ontology.

[0144] Next, normalization processing is performed, including type checking, alias alignment, duplicate item merging, and confidence calculation. This checks whether each entity and relation conforms to the ontology, merges aliases pointing to the same object, removes completely duplicate entities or triples, and generates result confidence based on model confidence, example support, and schema consistency. For results that fail to meet output constraints, the system triggers retry, format repair, or manual review.

[0145] Finally, the results of the normalization process are assembled into a current threat knowledge subgraph. This subgraph uses normalized threat entities as nodes and semantic relationships such as use, instruction, exploitation, attack target, and location as edges, providing a unified structure for subsequent attack chain and attack technique analysis.

[0146] For example, if the query text describes an attack organization using a loader to deliver ransomware, exploiting public vulnerabilities, and targeting computing infrastructure in a specific region, the system can construct the usage relationship between the attack organization and the loader, the instruction relationship between the loader and the ransomware, the exploitation relationship between the attack organization and the vulnerability, and the attack target relationship between the attack organization and the target asset.

[0147] In this embodiment, unstructured threat text is converted into a pattern-consistent graph structure before entering multi-agent reasoning, reducing the impact of entity type confusion, relation ambiguity, and illegal output on subsequent analysis.

[0148] In one feasible embodiment, step S400 is described in detail as follows:

[0149] Based on the current threat knowledge subgraph, relevant knowledge is retrieved from heterogeneous security knowledge sources such as threat intelligence knowledge graph, standardized attack tactics and techniques knowledge base, historical attack case library, evidence library, analysis rule library, computing power node asset ledger, node connection relationship and scheduling control plane log.

[0150] Specifically, the knowledge recall controller uses entities, relationships, and adjacency structures in the current threat knowledge subgraph as query conditions to retrieve related knowledge from the aforementioned heterogeneous security knowledge sources.

[0151] For each candidate knowledge item, the correlation degree between the candidate knowledge item and the current threat knowledge subgraph is calculated, and the recall knowledge set is obtained according to the following formula:

[0152] ;

[0153] in, This represents the set of knowledge to be recalled; This indicates the degree of association between the current threat knowledge subgraph and candidate knowledge items. This indicates selecting the top with the highest relevance. One candidate knowledge item; Represents a knowledge subgraph of current threats; Indicates candidate knowledge items; This indicates a heterogeneous source of security knowledge.

[0154] Based on this formula, select the top [top [topic] with the highest correlation]. A set of candidate knowledge items is obtained, and the recall knowledge set is obtained. During the recall process, the source, time, identifier, and association score of each knowledge item are recorded.

[0155] Subsequently, the recalled knowledge set and the current threat knowledge subgraph are written together into a shared inference storage area, which is a structured storage space for recording facts, evidence, subgraphs, task states, intermediate conclusions, conflict items, and adjudication results during multi-agent collaborative analysis.

[0156] In one feasible embodiment, step S500 is described in detail as follows:

[0157] Combination Figure 5 As shown, this embodiment provides a multi-agent collaborative method for performing attack chain reconstruction, attack technology mapping, attack organization profiling, evidence verification, and interpretation on the current threat knowledge subgraph.

[0158] The current threat knowledge subgraph and the recalled knowledge set are written into the shared inference storage area, and fact area, evidence area, subgraph area, intermediate conclusion area, conflict area, and adjudication result area are established. Each dedicated agent can only write results with task identifier, source identifier, and confidence level into the corresponding area.

[0159] The coordinating agent reads the shared inference storage area, generates a task plan based on the current threat knowledge subgraph, the recalled knowledge set, and the task routing rules, and calls multiple dedicated agents according to the task objectives and dependencies.

[0160] The dedicated intelligent agents used in this embodiment include an attack organization profiling intelligent agent, an attack technique mapping intelligent agent, an evidence intelligent agent, a reasoning intelligent agent, and an explanation intelligent agent.

[0161] The coordinating agent, combining a skill library and tool interfaces, routes various specialized agents; the attack profiling agent analyzes attackers and their activities; the attack technique mapping agent aligns and standardizes attack tactics and techniques; the evidence agent checks the support and consistency of conclusions; the reasoning agent expands the subgraph and reconstructs the attack chain; and the interpretation agent generates evidentiary explanations and Q&As. The adjudicating agent maintains a shared inference storage area and uniformly verifies the results of each agent.

[0162] The coordinating agent assigns tasks such as attack chain reconstruction, attack technique mapping, attack organization profiling, evidence verification, and interpretation generation to these specialized agents.

[0163] Each dedicated intelligent agent can perform graph query, knowledge retrieval, relation expansion, rule verification, and text generation through a preset skill library and tool interface, and write the intermediate results to the shared inference storage area.

[0164] The adjudicating agent maintains facts, evidence, subgraphs, and intermediate results in the shared inference store.

[0165] For inconsistent conclusions output by different agents, the adjudicating agent compares the consistency between the conclusion and the evidence in the current threat knowledge subgraph, the recall knowledge set, and the preset analysis rules, and performs retention, correction, rerouting, or rejection according to the sufficiency of the evidence.

[0166] The adjudicating agent will only include a candidate attack chain node, attack technique map, or attack organization attribute in the final conclusion if it is supported by facts in the current threat knowledge subgraph, evidence in the recalled knowledge set, or both.

[0167] If the technology number provided by the attack technology mapping agent is inconsistent with the verification result of the evidence agent, the adjudication agent reads the facts and evidence cited by both parties and prioritizes retaining conclusions that can be clearly supported by the original intelligence relationship, standardized knowledge items, or historical cases.

[0168] If it is still impossible to determine, the conclusion is marked as pending review, and the coordinating agent is required to recall the knowledge set or adjust the reasoning task.

[0169] In the final conclusion, each node and edge in the attack chain, the attack technique mapping, and the attack organization attribute are associated with at least one fact identifier or evidence identifier.

[0170] In one feasible embodiment, step S600 is described in detail as follows:

[0171] Based on the previous step, the adjudicating agent performs evidence consistency verification and conflict resolution on the intermediate conclusions of each dedicated agent in the shared inference storage area, resulting in the adjudication outcome.

[0172] The interpreting agent generates a threat analysis report based on the ruling, and can answer questions such as who the attacker is, how the attack was delivered, what vulnerabilities were exploited, what the target was, and what attack techniques were involved.

[0173] This invention also includes feedback and updates:

[0174] The conclusions of manual review can be written back to the historical case database or the analysis rule database. The system records the agent outputs that are accepted, corrected, and rejected, providing a basis for subsequent adjustments to task routing, evidence thresholds, and knowledge retrieval ranking; unreviewed generated content is not directly written into executable disposal strategies.

[0175] Through the above embodiments, multiple dedicated intelligent agents collaborate on a unified state of facts and evidence, and the adjudicating intelligent agent implements evidence gating to avoid conclusion drift and contradictions caused by independent reasoning of each dedicated intelligent agent.

[0176] In one feasible embodiment, the present invention provides a dual-channel threat retrieval and multi-agent collaborative analysis system for computing power network hubs, comprising:

[0177] The data access module is used to acquire threat intelligence and asset context to be analyzed from the computing power network hub;

[0178] A dual-channel retrieval module is used to generate a candidate pool through an entity-oriented sparse retrieval channel and a relationship-oriented dense retrieval channel;

[0179] The example selection module is used to select a set of context examples based on the dual-channel relevance combined with the maximum marginal relevance score.

[0180] The joint extraction module is used to construct contextual learning prompts and jointly extract threat entities and semantic relations under pattern constraints;

[0181] The knowledge graph module is used to normalize the extracted results and construct a knowledge subgraph of the current threat.

[0182] The knowledge retrieval module is used to retrieve knowledge from heterogeneous security knowledge sources and obtain a set of retrieved knowledge.

[0183] A multi-agent collaboration module is used to allocate analysis tasks to multiple specialized agents by coordinating agents;

[0184] The adjudication and shared storage module is used to maintain the shared inference storage area and perform evidence consistency verification and conflict resolution;

[0185] The analysis output module is used to output evidence-supported attack chains, attack technique mappings, attack organization profiles, and threat analysis reports.

[0186] The system can be deployed on the hub security operation platform, or the retrieval service, extraction service and multi-agent service can be deployed on the central computing node and regional collaboration node respectively.

[0187] The data access module receives secure data from computing centers, edge nodes, inter-node links, and scheduling control planes via message queues, interface services, or log collection components. The system forms logically isolated analysis tasks based on tenant identifiers, asset identifiers, and event identifiers, and appends asset importance, network location, and node relationships as asset context to the query text.

[0188] The dual-channel retrieval module and example selection module can operate as low-latency services. A sparse index stores threat entity terms, while a dense vector index stores security semantic representations; both retrieval methods are executed in parallel when threat intelligence arrives. For high-priority events, the number of deep sparse candidate sets, dense candidate sets, or context examples can be increased; for resource-constrained nodes, a smaller retrieval depth can be used, and multi-agent deep analysis can be delegated to the central node.

[0189] The system associates the current threat knowledge subgraph output by the joint extraction module with the computing power asset graph. If a threat entity contains vulnerabilities, target organizations, or management interfaces, the system locates potentially affected computing power nodes, scheduling components, and related services based on the asset ledger and node connection relationships, and writes the association results to the shared inference storage area.

[0190] The multi-agent collaboration module service allocates execution resources based on event priority, available computing power, and task dependencies. Evidence verification and adjudication tasks take precedence over interpretation and generation tasks; when evidence is insufficient, the system first outputs the flags to be reviewed and the confirmed facts, without generating a definitive attack attribution.

[0191] The final threat analysis report may include a summary of the original intelligence, normalized entities, relation triples, target computing power assets, candidate attack chains, confirmed attack techniques, attack organization profiles, evidence sources, confidence levels, and remediation recommendations. Remediation recommendations may include at least one of the following: investigation of related nodes, hardening of management interfaces, vulnerability patching, credential rotation, access restrictions, updating detection rules, and manual review.

[0192] Through this embodiment, the computing power network hub can convert unstructured threat intelligence across nodes and platforms into evidence-based analysis results associated with specific computing power assets, and allocate retrieval and inference tasks between the hub center and regional nodes according to resource conditions.

[0193] The technical terms, principles, or means related to the technical solutions of the present invention mentioned in the above embodiments, which are not described in detail above, are all well-known technologies or common practices that are known to those skilled in the art.

[0194] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs, characterized in that, Includes the following steps: Acquire threat intelligence and asset context to be analyzed from the computing power network hub, identify the threat intelligence to be analyzed as query text, and obtain a unified example library consisting of annotated threat intelligence examples; The corresponding candidate example sets are obtained from the unified example library through the entity-oriented sparse retrieval channel and the relation-oriented dense retrieval channel respectively, and the union of these sets is used to obtain the candidate pool. By integrating dual-channel relevance and combining the maximum marginal relevance to select scores and suppress example redundancy, a set of context examples is selected from the candidate pool, and a current threat knowledge subgraph is constructed based on the joint extraction of entities and semantic relationships under the constraints of the context example set and the preset threat intelligence ontology execution mode. Based on the current threat knowledge subgraph, a recall knowledge set is obtained from heterogeneous security knowledge sources, and the recall knowledge set and the current threat knowledge subgraph are written into the shared inference storage area. The coordinating agent invokes each dedicated agent based on the current threat knowledge subgraph, the recalled knowledge set, and the preset task routing rules; the adjudicating agent performs evidence consistency verification and conflict resolution of the intermediate conclusions of each dedicated agent based on the shared inference storage area. Based on the evidence consistency verification results, the interpreting agent outputs the attack chain, attack technology mapping, attack organization profile, and threat analysis report supported by the evidence.

2. The dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs according to claim 1, characterized in that, The entity-oriented sparse retrieval channel calculates sparse relevance using a term frequency-inverse document frequency sorting method: ; in, Indicates sparse correlation; Indicates the query text; Indicates candidate examples; Indicates terms; Indicates the inverse document frequency of a term; Indicates the frequency of a term in the candidate examples; This represents the preset parameters for controlling word frequency saturation. This represents the preset parameters for controlling length normalization. This represents the average length of examples in the Uniform Example Library; Select the top results based on sparsity correlation, from highest to lowest. From 10 candidate examples, a sparse candidate set is obtained; The relation-oriented dense retrieval channel uses a security domain text encoding model to generate query vectors and candidate vectors, and calculates dense relevance based on cosine similarity. Among them, the text encoding model in the security domain adopts an encoding model pre-trained or adapted with a cybersecurity corpus; and the top models are selected according to the density of relevance from high to low. 10 candidate examples are used to obtain a dense candidate set.

3. The dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs according to claim 1, characterized in that, The steps for selecting a set of context examples from the candidate pool are as follows: Normalize the sparse relevance and dense relevance in the candidate pool; for candidate examples and selected context example sets, calculate the redundancy of the candidate example relative to the selected context example set. When the selected context example set is empty, the redundancy is zero; when the selected context example set is not empty, the redundancy is the maximum value of the secure semantic similarity between the candidate example and any selected example in the selected context example set. The maximum marginal relevance selection score is calculated by combining redundancy: ; in, This represents the score for the selection of the maximum marginal relevance. Indicates the query text; Indicates candidate examples; This indicates the selected context example collection; Weighting coefficients representing the degree of sparse correlation; Weighting coefficients representing the degree of close correlation; Indicates the redundancy penalty weight; Indicates redundancy; This represents the average length of examples in the Uniform Example Library; In each round, candidate examples whose scores meet the preset conditions are added to the selected context example set, until a preset number of context examples are obtained or until the candidate pool is empty.

4. The dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs according to claim 1, characterized in that, The steps for constructing the current threat knowledge subgraph are as follows: The task instructions, preset threat intelligence ontology, structured output constraints, context example set and query text are combined into context learning prompts based on a preset form. The prompts are then input into a language model adapted to the security domain. Threat entities and semantic relationships are jointly extracted under the pattern constraints of the preset threat intelligence ontology. The extraction results are then subjected to type checking, alias alignment, duplicate item merging and confidence calculation, and finally assembled into the current threat knowledge subgraph. The language model adapted for the security domain can learn threat entity boundaries, relational semantics, and triple patterns by adopting a low-rank adaptation method while keeping the pre-trained parameters frozen. The entity types defined by the preset threat intelligence ontology include at least attackers, tools, malware, vulnerabilities, targets, and regions, and the relationship types include at least use, instruction, exploitation, attack target, and location. The structured output constraint limits the output to machine-readable data containing triples of entity, entity type, and semantic relation.

5. The dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs according to claim 1, characterized in that, The execution steps for obtaining the recall knowledge set from heterogeneous security knowledge sources are as follows: The heterogeneous security knowledge sources on which the recall knowledge set is based include at least one of the following: threat intelligence knowledge graph, standardized attack tactics and techniques knowledge base, historical attack case library, evidence library, analysis rule library, computing power node asset ledger, node connection relationship and scheduling control plane log; The knowledge recall controller uses entities, semantic relationships, and adjacency structures in the current threat knowledge subgraph as query conditions; Calculate the correlation degree between candidate knowledge items and the current threat knowledge subgraph, sort them from high to low, select the top preset number of candidate knowledge items that meet the preset correlation degree conditions from heterogeneous security knowledge sources, and record the source, time, identifier and correlation score of the candidate knowledge items to obtain the recall knowledge set.

6. The dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs according to claim 1, characterized in that, The coordinating agent reads the shared inference storage area and calls the attack organization profiling agent, attack technology mapping agent, evidence agent, inference agent, and explanation agent according to the task objectives and dependencies. Each specialized intelligent agent can perform graph query, knowledge retrieval, relation expansion, rule verification, and text generation through a pre-set skill library and tool interface; The adjudicating agent maintains facts, evidence, subgraphs, and intermediate results in the shared inference storage area. For inconsistent conclusions output by different agents, the adjudicating agent compares the consistency between the conclusion and the evidence in the current threat knowledge subgraph, the recall knowledge set, and the preset analysis rules, and performs retention, correction, rerouting, or rejection according to the sufficiency of the evidence.

7. The dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs according to claim 1, characterized in that, When performing evidence consistency verification, the adjudicating agent will only write a candidate attack chain node, attack technique mapping, or attack organization attribute into the evidence consistency verification result if it is supported by facts in the current threat knowledge subgraph, evidence in the recall knowledge set, or both. The method also includes writing the intermediate conclusions of each dedicated agent that have been manually reviewed back to the historical case library or analysis rule library, and recording the dedicated agent outputs that have been accepted, corrected or rejected. The content generated by each dedicated intelligent agent is not directly written into the executable disposal strategy without being verified for evidence consistency by the adjudicating intelligent agent or manually reviewed.

8. A dual-channel threat retrieval and multi-agent collaborative analysis system for implementing the method of any one of claims 1 to 7, characterized in that, include: The data access module is used to acquire threat intelligence and asset context to be analyzed from the computing power network hub. A dual-channel retrieval module is used to generate a candidate pool through an entity-oriented sparse retrieval channel and a relationship-oriented dense retrieval channel; The example selection module is used to select a set of context examples based on the dual-channel relevance combined with the maximum marginal relevance score. The joint extraction module is used to construct contextual learning prompts and jointly extract threat entities and semantic relations under pattern constraints; The knowledge graph module is used to normalize the extracted results and construct a knowledge subgraph of the current threat. The knowledge retrieval module is used to retrieve knowledge from heterogeneous security knowledge sources and obtain a set of retrieved knowledge. A multi-agent collaboration module is used to allocate analysis tasks to multiple specialized agents by coordinating agents; The adjudication and shared storage module is used to maintain the shared inference storage area and perform evidence consistency verification and conflict resolution; The analysis output module is used to output evidence-supported attack chains, attack technique mappings, attack organization profiles, and threat analysis reports.

9. An electronic device, comprising: The processor, communication interface, memory, and communication bus are connected, with the processor, communication interface, and memory communicating with each other via the communication bus. The feature is that the memory stores a computer program, which, when executed by the processor, causes the processor to perform the dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed, the computer program implements the dual-channel threat retrieval and multi-agent collaborative analysis method for computing power hubs as described in any one of claims 1-7.