Platform Massive Security Data Analysis Method and System Based on Large Model

By constructing a cross-domain security knowledge graph and generating causal chains through large-scale model reasoning, and combining incremental learning with online learning, the problem of insufficient interpretability in large-scale model security data analysis is solved, achieving complete security event analysis and efficient threat identification.

CN123116277APending Publication Date: 2026-11-17ZHONGAN ZHISHANG (BEIJING) DIGITAL TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610193821.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-02-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, security data analysis based on large models with hundreds of billions of parameters lacks interpretability analysis of attack paths, related entities, and causal logic, resulting in reduced analysis completeness.

Method used

By constructing a cross-domain security knowledge graph, combining it with large-scale model reasoning to generate causal chains of security events, and introducing incremental learning and online learning mechanisms, parallel processing of multi-domain data and consistency verification of interpretation are carried out, outputting feature-level, causal-level, and decision-level interpretations.

Benefits of technology

It combines multi-domain knowledge and causal reasoning of security incidents, enhancing the completeness and interpretability of security incident analysis, and improving the accuracy of threat identification and system efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116277A_ABST
    Figure CN123116277A_ABST
Patent Text Reader

Abstract

This invention relates to the field of security data analysis technology, and discloses a platform-based method and system for massive security data analysis based on a large model. The method includes: S1, data acquisition and semantic representation, collecting security data from multiple domains, standardizing the data format and extracting features, mapping it to a unified semantic vector space for semantic compression; S2, risk stratification and hierarchical processing; S3, construction of a cross-domain security knowledge graph; S4, generation and explanation output of security event causal chains; S5, incremental learning and online learning updates; S6, parallel and streaming processing of multi-domain data; and S7, explanation consistency verification and anomaly feedback. By constructing a cross-domain security knowledge graph through entity recognition and relation extraction, and aligning and completing it with the reasoning results of the large model, a security event causal chain is formed, outputting feature-level, causal-level, and decision-level explanations. This achieves the combination of multi-domain knowledge and causal reasoning, enhancing the completeness and interpretability of security event analysis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of security data analysis technology, specifically to a method and system for analyzing massive security data on a platform based on a large model. Background Technology

[0002] This stems from the groundbreaking progress made by large-scale models with hundreds of billions of parameters, such as GPT-4 and Gemini, in natural language processing and cross-modal understanding. These models have significantly improved their ability to understand complex semantic relationships through large-scale pre-training and fine-tuning techniques. At the same time, modern security platforms need to process massive amounts of security data from multiple sources and heterogeneous sources, such as network traffic, terminal logs, and threat intelligence. This data is characterized by its large scale, high dimensionality, and strong real-time requirements. Traditional analysis methods mainly use rule-based feature matching to output risk probability values ​​as analysis results.

[0003] However, current technology only outputs an abstract risk probability value, lacking interpretable analysis of attack paths, related entities, and causal logic, which leads to a reduction in the completeness of the analysis. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a platform-based massive security data analysis method and system based on large models, solving the problem of reduced integrity of component analysis due to the lack of interpretable analysis of attack paths, related entities, and causal logic.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a platform-based massive security data analysis method based on a large model, comprising: S1. Data Acquisition and Semantic Representation: Acquire security data from multiple fields, standardize the data format and extract features, input the data into a large model embedding algorithm, and map it to a unified semantic vector space for semantic compression. S2. Risk stratification and hierarchical processing: The semantically vectorized multi-domain security data is stratified and hierarchically processed to distinguish data of different risk levels or categories and to determine the corresponding subsequent analysis path. S3. Construction of cross-domain security knowledge graph: Based on multi-domain security data, a cross-domain security knowledge graph is constructed to store security-related entities and their causal relationships. S4. Generation and Explanation Output of Causal Chains for Security Events: Combining large-scale model reasoning and knowledge graphs, generate causal chains for security events and output explanation results. S5. Incremental learning and online learning updates: The large model and knowledge graph are updated through incremental learning and online learning mechanisms. Knowledge distillation is used to retain existing knowledge, and human feedback is combined to update model parameters. S6, Multi-domain Data Parallelism and Streaming Processing: Employs parallel computing, streaming processing, caching, and scheduling mechanisms to process multi-domain security data; S7, Interpretation Consistency Verification and Anomaly Feedback: Perform consistency verification and rationality verification on the causal chain and interpretation results, identify abnormal inputs, and feed the results back to S5.

[0006] Preferably, the data acquisition and semantic representation include: Security data from different fields is collected through log collection interfaces, network traffic collection probes, and behavior monitoring components. The data is format-converted using data parsing rules, and field information used to characterize the features of security events is extracted; Field information is input into the embedding algorithm, heterogeneous data is mapped to a unified semantic vector space, and the representation dimension is reduced by vector compression method.

[0007] Preferably, the risk stratification and classification process includes: Based on the risk assessment model, risk values ​​are calculated for the semantically vectorized multi-domain security data; When the risk value is below a preset threshold range, a lightweight analysis method is adopted; When the risk value is higher than the preset threshold range, an analysis method based on large model reasoning is adopted.

[0008] Preferably, the construction of the cross-domain security knowledge graph includes: Entity identification is performed on security data from multiple domains, where entities are objects related to security incidents; The relationships between the entities are identified based on the relationship extraction method; Store entities and their relationships in a cross-domain security knowledge graph.

[0009] Preferably, the security event causal chain generation and interpretation output includes: Generate causal chains between security events based on a large model; Align and complete the causal relationship chain with the cross-domain security knowledge graph; The output includes explanations at different levels, which can be presented in natural language or through visualization.

[0010] Preferably, the incremental learning and online learning updates include: New security data is processed in real time and semantic representations are generated. The knowledge distillation method updates model parameters while preserving existing knowledge. The model is optimized based on manual annotation or operation and maintenance feedback.

[0011] Preferably, the multi-domain data parallel and streaming processing includes: Multi-domain security data is fragmented and processed in parallel across multiple computing units; New security data is processed in real time; Tasks are assigned according to risk level, and historical data patterns are used for comparison.

[0012] Preferably, the consistency check includes: Explanatory information is obtained based on various explanation generation methods, including methods based on feature importance and methods based on causal reasoning; The results obtained from different interpretation methods are compared, and the comparison results are verified for reasonableness against a preset security knowledge base.

[0013] Preferably, the anomaly feedback includes: When the causal chain or explanation result does not meet the consistency or reasonableness conditions, the input is marked as an anomalous sample; Abnormal samples are fed back into incremental learning and online learning for subsequent training and updates, based on model updates and knowledge graph corrections.

[0014] A platform-based massive security data analysis system based on a large model, the system comprising: The data acquisition and semantic representation module is used to collect security data from multiple fields, standardize the data format and extract features, and map the data to a unified semantic vector space through an embedding algorithm. The risk stratification and graded processing module is used to perform risk assessment on semantically vectorized multi-domain security data, and classify the data into different levels based on the assessment results to determine the processing method. The cross-domain security knowledge graph construction module is used to perform entity recognition and relationship extraction on security data from multiple domains, and store the obtained security-related entities and their relationships as a cross-domain security knowledge graph. The security incident causal chain generation and explanation module is used to combine large model reasoning and knowledge graph to generate security incident causal chains and output multi-level explanation information; The incremental learning and online learning update module is used to update large models and knowledge graphs based on incremental learning and online learning mechanisms; The multi-domain data parallel and streaming processing module is used to process multi-domain security data through parallel computing, streaming computing, and cache scheduling. The consistency verification and anomaly feedback module is used to verify the consistency and rationality of the causal chain and the interpretation results, identify abnormal inputs, and feed the abnormal inputs back to the incremental learning and online learning update module.

[0015] This invention provides a method and system for analyzing massive amounts of platform security data based on a large model. It has the following beneficial effects: 1. This invention constructs a cross-domain security knowledge graph through entity recognition and relation extraction, and aligns and completes it with the reasoning results of large models to form a causal chain of security events, outputting feature-level, causal-level, and decision-level explanations, realizing the combination of multi-domain knowledge and causal reasoning, and enhancing the completeness and interpretability of security event analysis.

[0016] 2. This invention introduces a risk assessment model to calculate risk values ​​for multi-domain security data after semantic vectorization. Based on the risk level, it automatically selects lightweight analysis or large model inference to achieve dynamic allocation of computing resources. High-risk events are analyzed in depth, and low-risk events are processed quickly, thereby improving the overall efficiency of the system and the accuracy of threat identification.

[0017] 3. This invention collects security data from multiple domains, processes it through parsing and standardization, and then inputs it into an embedding algorithm to achieve semantic vectorization and compression of cross-domain data. This ensures the fusion and efficient processing of heterogeneous data in a unified semantic space, providing a consistent foundation for subsequent analysis.

[0018] 4. This invention combines incremental learning, online learning, and knowledge distillation mechanisms to absorb new data in real time while maintaining existing knowledge, and also introduces human feedback to optimize the model. Inconsistent or unreasonable interpretations are marked as anomalies and fed back to the learning module, forming a closed-loop update. Attached Figure Description

[0019] Figure 1 The flowchart shows the platform massive security data analysis method based on a large model according to the present invention. Figure 2 This is an architecture diagram of the platform-based massive security data analysis system based on a large model, as presented in this invention. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see the appendix Figure 1 This invention provides a method for analyzing massive security data on a platform based on a large model, including: S1. Data Acquisition and Semantic Representation: Acquire security data from multiple fields, standardize the data format and extract features, input the data into a large model embedding algorithm, and map it to a unified semantic vector space for semantic compression. Data acquisition and semantic representation include: Security data from different fields is collected through log collection interfaces, network traffic collection probes, and behavior monitoring components. The data is formatted using data parsing rules, and field information used to characterize security events is extracted. Field information is input into the embedding algorithm, heterogeneous data is mapped to a unified semantic vector space, and the representation dimension is reduced by vector compression method.

[0022] Specifically, security data from different fields is collected through log collection interfaces, network traffic collection probes, and behavior monitoring components. These fields include, but are not limited to, system logs, network traffic, application access records, user behavior data, and external threat intelligence. This enables real-time access to multi-source heterogeneous security data, ensuring the comprehensiveness and integrity of the data sources and providing a rich and continuous data foundation for subsequent analysis. The collected multi-domain data is then processed by data parsing rules and uniformly converted into a standardized format. During this process, field information that can characterize the features of security events is extracted, such as contextual attributes, behavioral patterns, or environmental characteristics. Through standardization and feature extraction, data from different formats and sources can be compared and processed under a unified data structure, improving the usability and comparability of the data and creating conditions for subsequent large-scale model analysis. Simultaneously, after standardization and feature extraction are completed, the feature data is input into a large model-based embedding algorithm for processing, enabling different types of data to be mapped to a unified semantic vector space. During the mapping process, vector compression methods are used to reduce the dimensionality of the high-dimensional representation, reducing redundant information and storage overhead, and transforming the originally heterogeneous data into a semantically consistent vector representation, thereby improving computational efficiency.

[0023] S2. Risk stratification and hierarchical processing: The semantically vectorized multi-domain security data is stratified and hierarchically processed to distinguish data of different risk levels or categories and to determine the corresponding subsequent analysis path. Risk stratification and classification include: Risk values ​​are calculated for semantically vectorized multi-domain security data based on a risk assessment model; When the risk value is below the preset threshold range, a lightweight analysis method is used; When the risk value exceeds the preset threshold range, an analysis method based on large model reasoning is adopted.

[0024] Specifically, the semantically vectorized multi-domain security data is input into the risk assessment model for processing. The risk assessment model can be implemented based on anomaly detection, classification algorithms, or clustering methods. The risk value corresponding to each data point is calculated. By quantifying the risk value, multi-domain security events that were originally difficult to compare directly can be mapped to a unified risk measurement system, thereby providing a basis for subsequent risk level classification and realizing the initial identification of potential security threats.

[0025] Furthermore, the semantic vectors obtained through the embedding algorithm are first input into a pre-trained risk assessment model, which can be implemented in the following way: Anomaly detection-based: A reference model is pre-built using normal data samples. When a new semantic vector is input, its deviation from the reference distribution is calculated. For example, the deviation value is mapped to a risk value by means of reconstruction error (such as the input-output difference of an autoencoder), anomaly score of an isolated forest, or the degree of deviation of confidence interval based on statistical features. Based on classification algorithms: Use labeled normal and abnormal samples to train a classification model, such as logistic regression, support vector machine or neural network. When new data is input, the model outputs the probability that it belongs to the "abnormal" category. This probability value or its function value can be used as the risk value. Clustering-based approach: By clustering historical data, normal behavior patterns and potential threat patterns are divided into different clusters. When new data arrives, the distance between it and the center of each cluster is calculated. If the distance to the smallest cluster exceeds a threshold, a higher risk value is assigned; otherwise, a lower risk value is assigned. When the risk value is below the preset threshold range, the data is classified as low-risk data and processed using a lightweight analysis method. The lightweight analysis method can include rule matching, pattern comparison or fast statistical methods, which can quickly complete the processing of low-risk data with low computational overhead, reduce the occupation of large model resources, and at the same time ensure the basic identification ability of low-risk events. Meanwhile, when the risk value exceeds the preset threshold range, the data is classified as high-risk data and is processed in depth using a large model-based reasoning analysis method. Combining contextual semantics and knowledge graph information, causal reasoning and pattern recognition are performed on the high-risk data to identify complex attack paths or potential threats, thereby improving the system's accuracy in identifying high-risk threats.

[0026] S3. Construction of cross-domain security knowledge graph: Based on multi-domain security data, a cross-domain security knowledge graph is constructed to store security-related entities and their causal relationships. The construction of cross-domain security knowledge graphs includes: Entity identification is performed on security data from multiple domains, where entities are objects related to security incidents; Identifying relationships between entities based on relation extraction methods; Store entities and their relationships in a cross-domain security knowledge graph.

[0027] Specifically, the first step is to perform entity recognition on security data from multiple domains. Entities are objects related to security events. By using named entity recognition methods based on large models or rule-based methods based on pattern matching, key objects such as accounts, processes, network nodes, service resources, or attack methods are extracted from log text, network traffic characteristics, or user behavior data. Through entity recognition, disordered security data is transformed into a structured set of objects, clarifying the core elements involved in security events and providing a foundation for subsequent relationship recognition and knowledge modeling. After entity recognition is completed, the interaction relationships between entities are identified through large model reasoning, such as the calling relationship between users and processes, the access relationship between processes and files, and the communication relationship between IP and services. This reveals the semantic associations of different objects in security incidents, establishes causal or interaction links between entities, and lays the foundation for the structured representation of knowledge graphs. Based on entities and their relationships, a unified cross-domain security knowledge graph is constructed by using a graph database, with entities as nodes and relationships as edges, to support querying, updating, and reasoning. The process of storing information into the graph enables structured management of security-related information, transforming heterogeneous data from multiple domains into computable knowledge structures, realizing global modeling and semantic fusion of cross-domain security events, and providing support for subsequent causal chain generation and interpretation.

[0028] S4. Generation and Explanation Output of Causal Chains for Security Events: Combining large-scale model reasoning and knowledge graphs, generate causal chains for security events and output explanation results. The output of the security event causal chain generation and interpretation includes: Generate causal chains between security events based on a large model; Align and complete the causal relationship chain with the cross-domain security knowledge graph; The output includes explanations at different levels, which can be presented in natural language or through visualization.

[0029] Specifically, by reasoning through semantically vectorized data using large models, and combining the contextual semantics, temporal sequence, and logical dependencies of events, a causal chain between security events is generated. This connects scattered security events into a logically directional causal structure, clarifies the causal dependencies between events, and effectively forms a traceable attack path, making it easier to discover the development process of potential threats. After generating the causal relationship chain, the event nodes in the causal chain are compared with the entity nodes in the cross-domain security knowledge graph. Through semantic similarity calculation and graph structure matching, the causal chain is aligned with the associated information in the existing knowledge graph, missing links are filled in, and existing knowledge is used to improve the completeness and consistency of the causal chain, ensuring that the generated attack chain is more accurate and comprehensive. After the causal chain generation and knowledge graph completion are completed, the explanation results are output. The explanation content includes three levels: feature level, causal level, and decision level. The feature level explanation is used to explain the role of key input features in event reasoning. The causal level explanation is used to show the causal relationship between events. The decision level explanation presents the overall analysis conclusion in natural language or visualization, thereby improving the understandability of security analysis results, helping operation and maintenance personnel to quickly grasp the core and logical relationship of events, and improving response and handling efficiency.

[0030] S5. Incremental learning and online learning updates: The large model and knowledge graph are updated through incremental learning and online learning mechanisms. Knowledge distillation is used to retain existing knowledge, and human feedback is combined to update model parameters. Incremental learning and online learning updates include: New security data is processed in real time and semantic representations are generated. The knowledge distillation method updates model parameters while preserving existing knowledge. The model is optimized based on manual annotation or operation and maintenance feedback.

[0031] Specifically, during operation, it receives continuously generated security data, processes this new data in real time and generates semantic representations. After standardization and feature extraction, the new data is input into a large model embedding algorithm to obtain semantic vectors consistent with existing data, thereby achieving continuous updates and unified expression of dynamic data, and ensuring that the latest security information can be incorporated into the analysis in a timely manner. In the incremental learning process, the knowledge distillation method is used to update the model parameters while maintaining the existing knowledge. By introducing the output results of the historical model as a soft constraint and combining it with the training results of the new data, the new model can avoid forgetting the original pattern when learning new samples and maintain its ability to recognize existing attack features and security patterns. Operations and maintenance personnel are allowed to manually annotate or provide feedback on some of the analysis results. This feedback is incorporated into the model update process to adjust model parameters and optimize the relational structure of the knowledge graph. In other words, the introduction of manual feedback improves the model's accuracy in complex or boundary cases and enhances the system's adaptability to new threats and unknown attack scenarios.

[0032] S6, Multi-domain Data Parallelism and Streaming Processing: Employs parallel computing, streaming processing, caching, and scheduling mechanisms to process multi-domain security data; Multi-domain data parallelism and streaming processing include: Multi-domain security data is fragmented and processed in parallel across multiple computing units; New security data is processed in real time; Tasks are assigned according to risk level, and historical data patterns are used for comparison.

[0033] Specifically, the collected multi-domain security data is processed in shards, with each shard allocated to different computing units for parallel execution. Through a distributed parallel architecture, large-scale data is broken down into multiple smaller batches, with multiple processing nodes running simultaneously, significantly reducing overall processing time and ensuring efficient response to large-scale security data. Upon receiving new security data, a streaming computing framework is used for real-time processing. Data is parsed and analyzed immediately upon arrival, rather than waiting for batch accumulation, thus enabling immediate detection and response to security incidents. This approach improves the system's real-time performance in dynamic environments, allowing for rapid identification in the early stages of a threat.

[0034] Based on the risk assessment results, processing tasks are prioritized, high-risk data is assigned to higher-priority computing queues, and historical data patterns are called for comparison. Through caching and indexing mechanisms, historical event characteristics can be quickly retrieved and matched with current high-risk data to help identify potential attack chains or recurring threat behaviors. This ensures that high-risk events are processed first, while also improving the overall accuracy and stability of detection.

[0035] S7, Interpretation Consistency Verification and Anomaly Feedback: Perform consistency verification and rationality verification on the causal chain and interpretation results, identify abnormal inputs, and feed the results back to S5.

[0036] Consistency verification includes: Explanatory information is obtained based on various explanation generation methods, including methods based on feature importance and methods based on causal reasoning; The results obtained from different interpretation methods are compared, and the comparison results are verified for reasonableness against a preset security knowledge base.

[0037] Specifically, explanatory information about causal chains and model inference results is obtained through various explanatory generation methods. These methods include feature importance-based methods and causal reasoning-based methods. The former generates explanatory information by analyzing the influence of input features on model decisions, while the latter generates causal explanations between events through a causal analysis framework. By combining different explanatory approaches, multi-dimensional explanatory results can be obtained, avoiding the limitations of a single method. After obtaining different interpretations, these results are compared for consistency to determine whether the conclusions of each interpretation method are consistent. When discrepancies arise between different interpretation methods, the system records and marks these discrepancies for further analysis. This comparison process ensures the consistency of the interpretation results across different dimensions, making the reasoning process more credible. After consistency comparison, the interpretation results are compared with a pre-set security knowledge base for rationality verification. Verification includes checking whether the event relationships in the causal chain conform to known security logic and whether the interpretation information is consistent with existing security patterns. When verification finds non-compliance, the system marks the input as an anomalous sample and feeds it back to the incremental learning and online learning modules for model updates and knowledge graph correction. This feedback mechanism ensures the system has self-correcting capabilities when facing anomalous inputs and improves overall robustness in subsequent iterations.

[0038] Furthermore, abnormal feedback includes: When the causal chain or explanation result does not meet the consistency or reasonableness conditions, the input is marked as an anomalous sample; Abnormal samples are fed back into incremental learning and online learning for subsequent training and updates, based on model updates and knowledge graph corrections.

[0039] Specifically, after verifying the consistency and rationality of the causal chain and the interpretation results, if any non-compliance is found, the corresponding input data is marked as an anomalous sample. The marking process is achieved by adding anomaly tags or anomaly identifier fields to the data records, so that such data can be distinguished and identified in subsequent processing. This ensures that anomalous inputs are not directly used for routine analysis and avoids unreliable information interfering with the system operation. Simultaneously, after labeling abnormal samples, these samples are passed to the incremental learning and online learning modules for subsequent training and updates. The abnormal samples are incorporated into the learning process, and by updating model parameters and correcting knowledge graph relationships, the model can better identify similar abnormal patterns during subsequent inference, thereby improving the model's adaptability and strengthening the knowledge graph's ability to describe new threat relationships. This gradually improves the system's robustness and accuracy through iteration.

[0040] Example: First, the data acquisition and semantic representation module connects to various databases, which cover multi-dimensional information such as politics, law, public security, public health, and natural disasters. After standardization and feature extraction, the collected data is mapped to a unified semantic vector space to achieve semantic compression and fusion of cross-domain information. The risk stratification and graded processing module calculates risk values ​​for data from different sources based on the risk assessment model. Low-risk information is quickly screened in a lightweight manner, while high-risk information is entered into deep reasoning analysis in a large model. The cross-domain security knowledge graph construction module identifies and stores core entities and their relationships in multiple databases to form a unified knowledge graph structure. Building upon this foundation, the security event causal chain generation and explanation module combines large-scale model reasoning and knowledge graphs to generate cross-domain causal chains. For example, it can demonstrate how political changes can lead to a deterioration in public security, which in turn triggers regional security threats. This chain is presented to users through multi-level explanations. Incremental learning and online learning mechanisms continuously absorb new data and human feedback, keeping the model and knowledge graph updated and evolving. The multi-domain data parallel and streaming processing module ensures the processing of data from multiple databases under high concurrency scenarios. Finally, the explanation consistency verification and anomaly feedback module compares and verifies the rationality of the causal chains and explanation results, feeding back illogical results to the learning module to further enhance the system's robustness.

[0041] Through the above embodiments, the system can achieve comprehensive analysis and causal modeling of information from multiple fields, including regional security threats, political instability, changes in laws and policies, terrorist activities, public security, public health, traffic safety, and natural disasters, forming a cross-domain, multi-dimensional, and interpretable security analysis closed loop. Please see the appendix Figure 2 A platform-based massive security data analysis system based on a large model, comprising: The data acquisition and semantic representation module is used to collect security data from multiple fields, standardize the data format and extract features, and map the data to a unified semantic vector space through an embedding algorithm. The risk stratification and graded processing module is used to perform risk assessment on semantically vectorized multi-domain security data, and classify the data into different levels based on the assessment results to determine the processing method. The cross-domain security knowledge graph construction module is used to perform entity recognition and relationship extraction on security data from multiple domains, and store the obtained security-related entities and their relationships as a cross-domain security knowledge graph. The security incident causal chain generation and explanation module is used to combine large model reasoning and knowledge graph to generate security incident causal chains and output multi-level explanation information; The incremental learning and online learning update module is used to update large models and knowledge graphs based on incremental learning and online learning mechanisms; The multi-domain data parallel and streaming processing module is used to process multi-domain security data through parallel computing, streaming computing, and cache scheduling. The consistency verification and anomaly feedback module is used to verify the consistency and rationality of the causal chain and the interpretation results, identify abnormal inputs, and feed the abnormal inputs back to the incremental learning and online learning update module.

[0042] Specifically, the data acquisition and semantic representation module acquires security data from different domains and performs standardized processing and semantic vector representation; the risk stratification and graded processing module assesses the risks of vectorized data and selects appropriate analysis methods according to the grade; the cross-domain security knowledge graph construction module forms a unified knowledge structure through entity recognition and relation extraction; the security event causal chain generation and explanation module generates causal chains by combining large-scale model reasoning and outputs multi-level explanation information; the incremental learning and online learning update module continuously optimizes the model and knowledge graph based on new data and human feedback; the multi-domain data parallel and streaming processing module is responsible for the efficient processing of large-scale and real-time data; and the explanation consistency verification and anomaly feedback module verifies the rationality of the causal chain results and feeds back abnormal samples to update the model.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing massive security data based on a large model platform, characterized in that, include: S1. Data Acquisition and Semantic Representation: Acquire security data from multiple fields, standardize the data format and extract features, input the data into a large model embedding algorithm, and map it to a unified semantic vector space for semantic compression. S2. Risk stratification and hierarchical processing: The semantically vectorized multi-domain security data is stratified and hierarchically processed to distinguish data of different risk levels or categories and to determine the corresponding subsequent analysis path. S3. Construction of cross-domain security knowledge graph: Based on multi-domain security data, a cross-domain security knowledge graph is constructed to store security-related entities and their causal relationships. S4. Generation and Explanation Output of Causal Chains for Security Events: Combining large-scale model reasoning and knowledge graphs, generate causal chains for security events and output explanation results. S5. Incremental learning and online learning updates: The large model and knowledge graph are updated through incremental learning and online learning mechanisms. Knowledge distillation is used to retain existing knowledge, and human feedback is combined to update model parameters. S6, Multi-domain Data Parallelism and Streaming Processing: Employs parallel computing, streaming processing, caching, and scheduling mechanisms to process multi-domain security data; S7, Interpretation Consistency Verification and Anomaly Feedback: Perform consistency verification and rationality verification on the causal chain and interpretation results, identify abnormal inputs, and feed the results back to S5.

2. The method of claim 1, wherein, The data acquisition and semantic representation include: Security data from different fields is collected through log collection interfaces, network traffic collection probes, and behavior monitoring components. The data is formatted using data parsing rules, and field information used to characterize security event features is extracted. Field information is input into an embedding algorithm to map heterogeneous data to a unified semantic vector space, and the representation dimension is reduced by vector compression methods. 3.The method of claim 1, wherein, The risk stratification and classification process includes: Risk values ​​are calculated for semantically vectorized multi-domain security data based on a risk assessment model; When the risk value is below a preset threshold range, a lightweight analysis method is adopted; When the risk value is higher than the preset threshold range, an analysis method based on large model reasoning is adopted.

4. The method of claim 1, wherein, The construction of the cross-domain security knowledge graph includes: Entity identification is performed on security data from multiple domains, where entities are objects related to security incidents; The relationships between the entities are identified based on the relationship extraction method; Store entities and their relationships in a cross-domain security knowledge graph.

5. The method of claim 1, wherein, The security event causal chain generation and explanation output includes: Generate causal chains between security events based on a large model; Align and complete the causal relationship chain with the cross-domain security knowledge graph; The output includes explanations at different levels, which can be presented in natural language or through visualization. 6.The method of claim 1, wherein, The incremental learning and online learning updates include: New security data is processed in real time and semantic representations are generated. The knowledge distillation method updates model parameters while preserving existing knowledge. The model is optimized based on manual annotation or operation and maintenance feedback.

7. The method of claim 1, wherein the method further comprises: The multi-domain data parallel and streaming processing includes: Multi-domain security data is fragmented and processed in parallel across multiple computing units; New security data is processed in real time; Tasks are assigned according to risk level, and historical data patterns are used for comparison.

8. The method for analyzing massive security data of a platform based on a large model according to claim 1, characterized in that, The consistency check includes: Explanatory information is obtained based on various explanation generation methods, including methods based on feature importance and methods based on causal reasoning; The results obtained from different interpretation methods are compared, and the comparison results are verified for reasonableness against a preset security knowledge base.

9. The method for analyzing massive security data of a platform based on a large model according to claim 1, characterized in that, The abnormal feedback includes: When the causal chain or explanation result does not meet the consistency or reasonableness conditions, the input is marked as an anomalous sample; Abnormal samples are fed back into incremental learning and online learning for subsequent training and updates, based on model updates and knowledge graph corrections.

10. A platform-based massive security data analysis system based on a large model, characterized in that: The system used for the platform massive security data analysis method based on a large model as described in any one of claims 1-9, the system comprising: The data acquisition and semantic representation module is used to collect security data from multiple fields, standardize the data format and extract features, and map the data to a unified semantic vector space through an embedding algorithm. The risk stratification and graded processing module is used to perform risk assessment on semantically vectorized multi-domain security data, and classify the data into different levels based on the assessment results to determine the processing method. The cross-domain security knowledge graph construction module is used to perform entity recognition and relationship extraction on security data from multiple domains, and store the obtained security-related entities and their relationships as a cross-domain security knowledge graph. The security incident causal chain generation and explanation module is used to combine large model reasoning and knowledge graph to generate security incident causal chains and output multi-level explanation information; The incremental learning and online learning update module is used to update large models and knowledge graphs based on incremental learning and online learning mechanisms; The multi-domain data parallel and streaming processing module is used to process multi-domain security data through parallel computing, streaming computing, and cache scheduling. The consistency verification and anomaly feedback module is used to verify the consistency and rationality of the causal chain and the interpretation results, identify abnormal inputs, and feed the abnormal inputs back to the incremental learning and online learning update module.