A network attack attribution analysis method and system integrating knowledge graphs and large models
By integrating knowledge graphs and large-scale models into a network attack attribution analysis method, this study solves the problems of rare path discovery and gang member identification in organized attacks within the network security protection system. This enables accurate attribution and prediction of network attacks, thereby improving network security and reliability.
Patent Information
- Application Number
- CN202611609786.6
- 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
When faced with organized and premeditated group-based cyberattacks, existing cybersecurity protection systems struggle to detect rare attack paths, find it difficult to correlate multi-source attacks, and address fragmented attack narratives, resulting in an inability to accurately identify attack groups and appropriate protection and response windows.
This study employs a network attack tracing and analysis method that integrates knowledge graphs and large models. By deconstructing paths, identifying the access status of multiple attack sources, constructing multi-dimensional knowledge graphs, and analyzing large language models, it reconstructs attack chains and predicts future attack characteristics.
It enables accurate detection of rare access behaviors, cross-IP and cross-account identification of attack group members, reveals hidden collaboration, automatically reconstructs attack narratives, supports proactive defense decisions, and improves network security and reliability.
Smart Images

Figure CN123123028A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network security, and in particular to a method and system for tracing and analyzing network attacks by integrating knowledge graphs and large models. Background Technology
[0002] Current cybersecurity protection systems in power companies include multiple layers of defense such as perimeter protection, host protection, and traffic monitoring. However, when facing organized and premeditated group cyberattacks, traditional single-point alarm analysis models have the following shortcomings: First, the concealment of attack paths leads to the failure of single-point detection. Attackers commonly use distributed jump servers, proxy IP rotation, and low-frequency, slow probing techniques, making it easy for a single access behavior to be submerged in normal business traffic. Traditional threshold- or signature-based detection engines struggle to detect rare probing paths pointing to sensitive interfaces. Second, it is difficult to correlate multi-source attacks and determine the group's composition. First, the attack group uses multiple IPs, different accounts, and heterogeneous tools to conduct coordinated infiltration of targets. Existing methods lack the ability to correlate cross-source behaviors and cannot penetrate proxy disguises to identify the same source manipulator behind the scenes, resulting in a situation where "you can only see the trees but not the forest." Second, the attack narrative is fragmented and lacks a global perspective. Disparate security alerts, logs, and intelligence data cannot be automatically woven into a coherent attack storyline, making it difficult to reconstruct the complete attack chain of "multi-point infiltration → coordinated exploitation → privilege diffusion." It is also impossible to distinguish the division of labor among different roles within the attack group, thus missing the best window for protection and response. Summary of the Invention
[0003] Given that existing network security protection measures have problems such as insufficient ability to detect rare attack paths, difficulty in associating multi-source attack groups, and reliance on human experience to reconstruct attack chains, they are unable to accurately and effectively respond to group attacks on the network, thus reducing the security and reliability of network operations.
[0004] To address the above problems, this invention provides a network attack attribution analysis method that integrates knowledge graphs and large-scale models, including: Step S1: Obtain multimodal network information, perform path deconstruction on the multimodal network information to obtain a complete access trajectory; determine rare access behaviors of the network based on the complete access trajectory; Step S2: Identify the access status of the rare access behavior from multiple attack sources and generate an attack source access vector; based on the cluster analysis of the attack source access vector, trace the attack groups of the network. Step S3: Construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph to determine potential attack penetration channels; construct a three-dimensional spatiotemporal attack network based on the potential attack penetration channels; Step S4: Perform large language model analysis on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative features of the network; predict the future attack features of the network based on the dynamic attack narrative features.
[0005] Optionally, in step S1, multimodal network information is acquired, and path deconstruction is performed on the multimodal network information to obtain a complete access trajectory; based on the complete access trajectory, rare access behaviors of the network are determined, including: Based on the distribution of network activity time intervals, multimodal network information is collected and preprocessed; wherein, the multimodal network information includes Web access logs, API call records, and directory traversal behavior records; The multimodal network information is deconstructed to obtain several path elements; wherein, the several path elements include protocol, domain name, interface level, operation type and behavioral parameter characteristics; based on the several path elements, a complete access trajectory is constructed; Perform global frequency statistics on all complete access trajectories, remove normal business paths, and obtain several preliminary complete access trajectories. Inverse document frequency quantization is performed on the several initial screening complete access trajectories to identify rare access behaviors of the network.
[0006] Optionally, in step S2, the rare access behavior is subjected to multi-attack source access state identification to generate an attack source access vector; based on the clustering analysis of the attack source access vector, the attack group of the network is traced, including: Extract multidimensional temporal features of the rare access behaviors; wherein, the multidimensional temporal features include access time interval, temporal variation law of behavioral parameters, and agent rotation time sequence; Based on the multi-dimensional temporal features, the access behavior of multiple attack sources is matched to obtain multi-dimensional access attribute elements of multiple attack sources; wherein, the multi-dimensional access attribute elements include the path rarity sequence of rare access behaviors initiated by multiple attack sources, access request parameters, access time distribution, access session duration, and access location jump. The multi-dimensional access attribute elements are subjected to semantic vector transformation and dimensionality reduction to obtain the attack source access vector. A multi-attack source similarity matrix is constructed by performing composite metric processing on the attack source access vector; wherein, the composite metric processing includes cosine similarity and Euclidean distance metrics; Based on the multi-attack source similarity matrix, the attack sources are clustered and merged to identify core gang clusters and peripheral interference sources; the core gang clusters are corrected based on the peripheral interference sources to determine the attack gang groups of the network.
[0007] Optionally, in step S3, a multidimensional knowledge graph of the attack group is constructed, and path retrieval is performed on the multidimensional knowledge graph to determine potential attack penetration channels; based on the potential attack penetration channels, a three-dimensional spatiotemporal attack network is constructed, including: Obtain the multi-dimensional entity attributes corresponding to the attack group; wherein, the multi-dimensional entity attributes include attack source identity attributes, target asset attributes, vulnerability attributes, and attack time attributes; import the multi-dimensional entity attributes into a graph engine to perform entity relationship definition and construct a multi-dimensional knowledge graph of the attack group; A path retrieval from the attack source to the target asset is performed on the multidimensional knowledge graph to obtain several feasible attack penetration channels; vulnerability remediation prediction is performed on each feasible attack penetration channel to determine whether each feasible attack penetration channel belongs to a potential attack penetration channel. By identifying shared resources and identifying the collaborative patterns of the attack groups among all potential attack channels, a three-dimensional spatiotemporal attack network can be constructed.
[0008] Optionally, in step S4, large language model analysis is performed on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative features of the network; based on the dynamic attack narrative features, the future attack characteristics of the network are predicted, including: The global arrangement of attack event timestamps and the division of roles among gang members in the three-dimensional spatiotemporal attack network are extracted using a large language model. Spatiotemporal reasoning is performed on the global arrangement of the attack event timestamps and the division of roles among the gang members to obtain the dynamic attack narrative features of the network; wherein, the dynamic attack narrative features include the temporal and spatial collaborative relationships of multi-point attack behaviors initiated by different gang members with different role labels. Based on the dynamic attack narrative characteristics and the spatiotemporal characteristics of the network's asset spatial layout and exploitable vulnerabilities, predict the target objects that the network will be attacked in the future and the types of attacks they will be subjected to. Based on the spatiotemporal characteristics of all future targets and the types of attacks they are subjected to, a comprehensive attack protection profile report for the network is generated.
[0009] As one aspect of the present invention, the present invention also provides a network attack attribution analysis system that integrates knowledge graphs and large models, including: The path deconstruction module is used to acquire multimodal network information, deconstruct the path of the multimodal network information, and obtain a complete access trajectory. The access behavior determination module is used to determine rare access behaviors of the network based on the complete access trajectory. The access vector generation module is used to identify the access status of the rare access behavior from multiple attack sources and generate an attack source access vector. The gang tracing module is used to trace the attack gangs of the network based on the clustering analysis of the access vectors of the attack sources; The attack channel determination module is used to construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph, and determine potential attack penetration channels. An attack network construction module is used to construct a three-dimensional spatiotemporal attack network based on the potential attack penetration channels. The attack narrative determination module is used to perform large language model analysis on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative characteristics of the network. An attack prediction module is used to predict the future attack characteristics of the network based on the dynamic attack narrative characteristics.
[0010] Optionally, the path deconstruction module is used to acquire multimodal network information, perform path deconstruction on the multimodal network information, and obtain a complete access trajectory, including: Based on the distribution of network activity time intervals, multimodal network information is collected and preprocessed; wherein, the multimodal network information includes Web access logs, API call records, and directory traversal behavior records; The multimodal network information is deconstructed to obtain several path elements; wherein, the several path elements include protocol, domain name, interface level, operation type and behavioral parameter characteristics; based on the several path elements, a complete access trajectory is constructed; The access behavior determination module is used to determine rare access behaviors of the network based on the complete access trajectory, including: Perform global frequency statistics on all complete access trajectories, remove normal business paths, and obtain several preliminary complete access trajectories. Inverse document frequency quantization is performed on the several initial screening complete access trajectories to identify rare access behaviors of the network.
[0011] Optionally, the access vector generation module is used to identify the access status of the rare access behavior from multiple attack sources and generate an attack source access vector, including: Extract multidimensional temporal features of the rare access behaviors; wherein, the multidimensional temporal features include access time interval, temporal variation law of behavioral parameters, and agent rotation time sequence; Based on the multi-dimensional temporal features, the access behavior of multiple attack sources is matched to obtain multi-dimensional access attribute elements of multiple attack sources; wherein, the multi-dimensional access attribute elements include the path rarity sequence of rare access behaviors initiated by multiple attack sources, access request parameters, access time distribution, access session duration, and access location jump. The multi-dimensional access attribute elements are subjected to semantic vector transformation and dimensionality reduction to obtain the attack source access vector. The gang tracing module is used to trace the attack gangs of the network based on cluster analysis of the attack source access vectors, including: A multi-attack source similarity matrix is constructed by performing composite metric processing on the attack source access vector; wherein, the composite metric processing includes cosine similarity and Euclidean distance metrics; Based on the multi-attack source similarity matrix, the attack sources are clustered and merged to identify core gang clusters and peripheral interference sources; the core gang clusters are corrected based on the peripheral interference sources to determine the attack gang groups of the network.
[0012] Optionally, the attack channel determination module is used to construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph, and determine potential attack penetration channels, including: Obtain the multi-dimensional entity attributes corresponding to the attack group; wherein, the multi-dimensional entity attributes include attack source identity attributes, target asset attributes, vulnerability attributes, and attack time attributes; import the multi-dimensional entity attributes into a graph engine to perform entity relationship definition and construct a multi-dimensional knowledge graph of the attack group; A path retrieval from the attack source to the target asset is performed on the multidimensional knowledge graph to obtain several feasible attack penetration channels; vulnerability remediation prediction is performed on each feasible attack penetration channel to determine whether each feasible attack penetration channel belongs to a potential attack penetration channel. The attack network construction module is used to construct a three-dimensional spatiotemporal attack network based on the potential attack penetration channels, including: By identifying shared resources and identifying the collaborative patterns of the attack groups among all potential attack channels, a three-dimensional spatiotemporal attack network can be constructed.
[0013] Optionally, the attack narrative determination module is used to perform large language model analysis on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative features of the network, including: The global arrangement of attack event timestamps and the division of roles among gang members in the three-dimensional spatiotemporal attack network are extracted using a large language model. Spatiotemporal reasoning is performed on the global arrangement of the attack event timestamps and the division of roles among the gang members to obtain the dynamic attack narrative features of the network; wherein, the dynamic attack narrative features include the temporal and spatial collaborative relationships of multi-point attack behaviors initiated by different gang members with different role labels. The attack prediction module is used to predict the future attack characteristics of the network based on the dynamic attack narrative characteristics, including: Based on the dynamic attack narrative characteristics and the spatiotemporal characteristics of the network's asset spatial layout and exploitable vulnerabilities, predict the target objects that the network will be attacked in the future and the types of attacks they will be subjected to. Based on the spatiotemporal characteristics of all future targets and the types of attacks they are subjected to, a comprehensive attack protection profile report for the network is generated.
[0014] The beneficial effects of the above-mentioned technical solutions provided in the embodiments of the present invention include at least the following: This invention provides a method and system for network attack tracing and analysis that integrates knowledge graphs and large-scale models. It acquires multimodal network information, performs path deconstruction on this information to obtain complete access trajectories, thereby identifying rare access behaviors. For these rare access behaviors, it identifies multi-attack source access states, generates attack source access vectors, and traces attack groups within the network through cluster analysis. It constructs a multidimensional knowledge graph of these attack groups, performs path retrieval on the graph to identify potential attack penetration channels, and thus constructs a three-dimensional spatiotemporal attack network. Finally, it performs large-scale language model analysis on the spatiotemporal attack network to obtain dynamic attack narrative features, thereby predicting future attack characteristics. In the face of group-based attack scenarios, this invention, based on rare access behavior detection, knowledge graph association modeling, and large-scale language model deduction, accurately reconstructs complete attack chains and distinguishes the roles and responsibilities of different groups, improving network security and reliability.
[0015] The network attack tracing and analysis method and system that integrates knowledge graphs and large models of the present invention have the following advantages: First, the present invention can accurately capture covert detections and effectively detect low-frequency slow attacks through rare access behavior detection, thereby reducing the false negative rate of attacks. Second, this invention can penetrate disguises to identify members of attack groups. Through semantic vector encoding and cluster analysis, it can merge attackers from the same source across IPs and accounts, thereby improving the accuracy of identifying the roles of attack group members. Third, this invention can uncover implicit collaborations through three-dimensional correlation, revealing indirect correlations that traditional correlation rules cannot discover by constructing a multi-dimensional knowledge graph, thus expanding the depth of attack threat detection; Fourth, this invention can achieve automatic narration and reduce cognitive load by transforming massive discrete data into structured data through a large language model, effectively shortening the time required for attack tracing. Fifth, this invention can support proactive defense decision-making, providing the defender with predictions of the attacker's next move through a global timeline, and supporting the formulation of pre-deployment strategies.
[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.
[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the network attack tracing and analysis method that integrates knowledge graphs and large models provided in this embodiment of the invention.
[0019] Figure 2 This is a schematic diagram of the network attack tracing and analysis system that integrates knowledge graphs and large models provided in this embodiment of the invention. Detailed Implementation
[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0021] Please see Figure 1 As shown in the figure, an embodiment of this application provides a network attack attribution analysis method that integrates knowledge graphs and large models. This network attack attribution analysis method that integrates knowledge graphs and large models includes: Step S1: Obtain multimodal network information, perform path deconstruction on the multimodal network information to obtain a complete access trajectory; determine rare access behaviors of the network based on the complete access trajectory; Step S2: Identify the access status of rare access behaviors from multiple attack sources and generate attack source access vectors; based on the cluster analysis of the attack source access vectors, trace the attack groups in the network. Step S3: Construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph to identify potential attack penetration channels; construct a three-dimensional spatiotemporal attack network based on the potential attack penetration channels; Step S4: Perform large language model analysis on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative characteristics of the network; based on the dynamic attack narrative characteristics, predict the future attack characteristics of the network.
[0022] This network attack attribution analysis method, which integrates knowledge graphs and large models, accurately reconstructs the complete attack chain and distinguishes the roles and responsibilities of different groups when facing gang-related attack scenarios. This is achieved through rare access behavior detection, knowledge graph association modeling, and large language model inference. This improves network security and reliability.
[0023] In another embodiment, in step S1, multimodal network information is acquired, and path deconstruction is performed on the multimodal network information to obtain a complete access trajectory; based on the complete access trajectory, rare access behaviors of the network are determined, including: Based on the distribution of network activity time intervals, multimodal network information is collected and preprocessed; among which, multimodal network information includes Web access logs, API call records, and directory traversal behavior records; The multimodal network information is deconstructed to obtain several path elements, including protocol, domain name, interface level, operation type, and behavioral parameter characteristics. Based on these path elements, a complete access trajectory is constructed. Perform global frequency statistics on all complete access trajectories, remove normal business paths, and obtain several preliminary complete access trajectories. Inverse document frequency quantization is performed on several initially screened complete access trajectories to identify rare access behaviors on the network.
[0024] The beneficial effects of the above embodiments are that group attacks on the network typically use normal network operations as cover. That is, they utilize the network's high-frequency, high-traffic normal business activities as a background, while group members, based on their own role labels, launch low-frequency access behaviors that are clearly different from normal business activities. Therefore, group attacks usually occur during periods of high network activity. Considering that group attacks have a higher probability of occurring during periods of high network activity, their probability of being detected is also higher. Furthermore, group attacks need to be separated from the massive amount of normal network activity, resulting in a high volume of business behavior data for analysis and processing. Performing business behavior data analysis and processing throughout the entire network lifecycle requires not only a large amount of readily available computing resources but also a considerable amount of time to wait for processing results, demanding significant investment in both computing resources and time costs. Considering the characteristic of organized attacks needing to disguise themselves as normal business operations, time-domain differentiated data collection and analysis can be performed on the network. This involves distinguishing between busy and idle time periods based on network activity levels, and then performing differentiated data collection during these periods. For example, data can be collected at a higher frequency during busy periods and at a lower frequency during idle periods. This allows for the allocation of computing resources to the busy periods and effectively reduces the total computing resources required throughout the network's lifecycle, enabling rational allocation of data collection and analysis decisions within limited computing resources. Furthermore, organized attacks leave traces in various forms during network operation. To comprehensively and accurately detect organized attacks, multimodal network information such as web access logs, API call records, and directory traversal behavior records is collected. This facilitates subsequent analysis of all business behaviors from multiple dimensions, including web access logs, API call records, and browser directory query behavior records.
[0025] It is understandable that the aforementioned multimodal network information corresponds to various types of business behaviors, each with a specific access path within the network. To characterize the network spatial layout of the access paths for each business behavior, the multimodal network information is deconstructed, extracting several path elements. Each path element represents the entity representing the access path within the network. These path elements may include, but are not limited to, protocols (i.e., the network protocols upon which the access path relies), domain names (i.e., the access domain names corresponding to the access path), interface levels (i.e., the API interface levels called by the access path), operation types (i.e., the type of access operation implemented by the access path), and behavioral parameter characteristics (such as the initiation time, duration, and gateway node locations traversed by the access path). Based on these path elements, a comprehensive reconstruction of the access path is performed, constructing a complete access trajectory. This complete access trajectory corresponds to a point-to-point complete access process path from the access initiator to the accessed object. During network operation, a large number of normal business processes are inevitably involved. The paths for these normal tasks may include, but are not limited to, homepage access and regular query interfaces. When a network is subjected to a coordinated attack, members of the group will mix a small number of malicious attacks into a large amount of normal business traffic. These malicious attacks typically converge on sensitive interfaces, which may include, but are not limited to, management backends, database management interfaces, and unauthorized API endpoints. In practice, a global frequency analysis can be performed on all complete access paths to determine the frequency of occurrence of each complete access path during the entire network operation. If the frequency is greater than or equal to a preset frequency threshold, its corresponding normal business path is determined; if the frequency is less than the preset frequency threshold, it is determined that it does not correspond to a normal business path. Simultaneously, the normal business paths are extracted, and the remaining complete access paths are used as initial screening complete access paths. Furthermore, an inverse document frequency (IDF) algorithm is introduced to quantify and determine the path sparseness (IDF) of each initially screened complete access path, where IDF(path) = log0 10 (N / (1+DF(path))), where N represents the total number of access logs for the initial screening of complete access trajectories, and DF(path) represents the number of independent sessions corresponding to the path of the initial screening complete access trajectory; the higher the value of the path rarity IDF(path), the higher the rarity of the path. By comparing the path rarity IDF(path) threshold of each initial screening complete access trajectory, rare access behaviors are identified, providing a basis for path tracing for subsequent identification of attack groups.
[0026] In another embodiment, in step S2, rare access behaviors are identified by multi-attack source access status to generate attack source access vectors; based on cluster analysis of the attack source access vectors, the network's attack groups are traced, including: Extract multidimensional temporal features of rare access behaviors; among which, multidimensional temporal features include access time interval, temporal variation pattern of behavioral parameters, and agent rotation time sequence; Based on multi-dimensional temporal features, the access behavior of multiple attack sources is matched to obtain multi-dimensional access attribute elements of multiple attack sources. Among them, the multi-dimensional access attribute elements include the path rarity sequence of rare access behaviors initiated by multiple attack sources, access request parameters, access time distribution, access session duration, and access location jump. Semantic vector transformation and dimensionality reduction are performed on multi-dimensional access attribute elements to obtain the attack source access vector; A multi-attack source similarity matrix is constructed by performing composite metric processing based on the attack source access vector; the composite metric processing includes cosine similarity and Euclidean distance metrics. Based on the multi-attack source similarity matrix, the attack sources are clustered and merged to identify the core gang clusters and peripheral interference sources; the core gang clusters are corrected based on the peripheral interference sources to determine the network's attack gang groups.
[0027] Group attacks involve multiple attack sources acting in different roles to launch different types of attacks on the network. If the identity of the attack source is inferred at the network level based on detected attack behaviors, different attack behaviors originating from the same source may be mistakenly identified as originating from multiple sources, leading to misidentification of the attack source. To accurately and uniquely identify the attack source at the attack initiation side, it is necessary to characterize the access behavior of multiple attack sources at the network level. It is understandable that there is a significant overlap between the access behavior of multiple attack sources and rare access behaviors. By extracting multi-dimensional temporal features from rare access behaviors, a consistent temporal feature representation of the access behavior of multiple attack sources can be achieved to a high degree. To accurately characterize the temporal variation of rare access behaviors, the following parameters are extracted: access time interval (i.e., the time interval between two adjacent rare access behaviors), temporal variation law of behavioral parameters (i.e., the change law of the operational parameters of rare access behaviors over time), and proxy rotation sequence (i.e., the rotation order of the proxy IPs applicable to rare access behaviors over time). Next, attack source access behavior matching and filtering are performed on the above multi-dimensional temporal features. This involves filtering multi-dimensional access attribute elements from the multi-dimensional temporal features to match attack source access behaviors. These multi-dimensional access attribute elements may include, but are not limited to, the path rarity sequence of rare access behaviors initiated by multiple attack sources (i.e., the path rarity value queue of all rare access behaviors), access request parameters (i.e., the information type and data volume requested by the rare access behavior), access time distribution (i.e., the occurrence time of the rare access behavior), access session duration (i.e., the duration of the rare access behavior), and access location jumps (i.e., the network node location jumps accessed by the rare access behavior). Semantic vector transformation and principal component analysis are then performed on the above multi-dimensional access attribute elements to reduce dimensionality, resulting in attack source access vectors. These low-dimensional dense vectors represent the attack source's access state to the network.
[0028] Furthermore, a composite metric combining cosine similarity and Euclidean distance is used to process the aforementioned attack source access vectors to construct a multi-attack source similarity matrix. This composite metric takes into account both the behavioral pattern similarity and attribute distance of multiple attack source accesses. Based on the multi-source attack similarity matrix, each attack source is clustered and merged using density clustering algorithms such as DBSCAN or HDBSCAN to identify core gang clusters and peripheral interference sources. The core gang cluster refers to a group of gang members that meet a preset clustering distance range, while the peripheral interference source refers to a member that has specific types of business interactions with the core gang cluster. The duration of specific types of business interactions between each peripheral interference source and the core gang cluster is obtained, and it is determined whether the duration exceeds a preset duration threshold. If so, the corresponding peripheral interference source is included in the core gang cluster; otherwise, it is not included. By correcting the core gang cluster in this way, it effectively penetrates camouflage methods such as proxy IPs, distributed jump servers, and cloud host rotation, accurately identifies members of the same attack gang, and determines the final attack gang group.
[0029] In another embodiment, in step S3, a multi-dimensional knowledge graph of the attack group is constructed, and path retrieval is performed on the multi-dimensional knowledge graph to determine potential attack penetration channels; based on the potential attack penetration channels, a three-dimensional spatiotemporal attack network is constructed, including: Obtain the multi-dimensional entity attributes corresponding to the attack group; among which, the multi-dimensional entity attributes include attack source identity attributes, target asset attributes, vulnerability attributes, and attack time attributes; import the multi-dimensional entity attributes into the graph engine to perform entity relationship definition and construct a multi-dimensional knowledge graph of the attack group; A path retrieval from the attack source to the target asset is performed on the multidimensional knowledge graph to obtain several feasible attack penetration channels; vulnerability remediation prediction is performed on each feasible attack penetration channel to determine whether each feasible attack penetration channel belongs to a potential attack penetration channel. By identifying shared resources and identifying the collaborative patterns of attack groups across all potential attack channels, a three-dimensional spatiotemporal attack network can be constructed.
[0030] Each member of the attack group has a specific role, which may include, but is not limited to, reconnaissance, exploitation, and lateral movement. Reconnaissance is responsible for port scanning, directory brute-force attacks, and information gathering; exploitation is responsible for vulnerability exploitation, payload delivery, and initial privilege acquisition; and lateral movement is responsible for internal network penetration, credential theft, and privilege maintenance. It is evident that the different entity attributes of each member of the attack group directly affect the attacks launched by the group within the network. To comprehensively characterize the spatiotemporal state of the attack group's attacks within the network, we first obtain multi-dimensional entity attributes corresponding to the attack source identity, target asset attributes, vulnerability attributes, and attack time attributes. Attack source identity attributes may include, but are not limited to, the attack source's IP address, domain name, account, and tool fingerprint; target asset attributes may include, but are not limited to, servers, databases, and business systems; vulnerability attributes may include, but are not limited to, CVE number, vulnerability type, and CVSS score; and attack time attributes may include, but are not limited to, attack stage, duration, and peak time. The aforementioned diverse entity attributes are imported into graph engines such as Neo4j to execute the relationships between entities, constructing a four-dimensional knowledge graph about attackers, assets, vulnerabilities, and time. These relationships may include, but are not limited to, the exploitation relationships of attacks on vulnerabilities, the targeting relationships of attacks on assets, the existence relationships of vulnerabilities within assets, the temporal relay relationships between different attacks, the sharing relationships of C2 domains, tool fingerprints, and digital certificates by different attacks, and the causal relationships between vulnerabilities and privilege acquisition / data leakage.
[0031] Community discovery algorithms, such as the Louvain algorithm, are used to identify closely cooperating attack groups. For example, path retrieval algorithms, such as shortest path or full path traversal, are used to perform path retrieval from the attack source to the target asset, resulting in several feasible attack penetration channels. Vulnerability remediation predictions are then performed on each feasible attack penetration channel to determine whether it belongs to a potential attack penetration channel. For instance, based on the predicted vulnerability remediation completion time associated with the feasible attack penetration channel, it is determined whether the effective time of the feasible attack penetration channel coincides with the predicted vulnerability remediation completion time. If so, the feasible attack penetration channel is determined not to be a potential attack penetration channel; otherwise, it is determined to be a potential attack penetration channel. Furthermore, Common Neighbors analysis is used to uncover the hidden associations of shared infrastructure among all potential attack penetration channels, transforming discrete clustering results into a three-dimensional spatiotemporal attack network. This reveals attack group collaboration patterns such as relay attacks from attack sources, shared vulnerability exploit payloads, and shared C2 infrastructure, providing a basis for generating dynamic attack narratives.
[0032] In another embodiment, in step S4, a large language model analysis is performed on the three-dimensional spatiotemporal attack network to obtain the network's dynamic attack narrative characteristics; based on the dynamic attack narrative characteristics, the network's future attack characteristics are predicted, including: The global arrangement of attack event timestamps and the division of roles among gang members in a three-dimensional spatiotemporal attack network are extracted using a large language model. Spatiotemporal reasoning is performed on the global arrangement of attack event timestamps and the division of roles among gang members to obtain the dynamic attack narrative characteristics of the network. Among them, the dynamic attack narrative characteristics include the temporal and spatial collaborative relationships of multi-point attack behaviors launched by different gang members with different role labels. Based on the dynamic attack narrative characteristics and the spatial layout of network assets and the spatiotemporal characteristics of exploitable vulnerabilities, predict the target objects of future attacks on the network and the types of attacks they will be subjected to. Based on the spatiotemporal characteristics of all future targets and the types of attacks they may be subjected to, a comprehensive attack and defense profile report of the network is generated.
[0033] In practice, a large language model is used to extract the global distribution of attack event timestamps and the division of roles among different groups in a three-dimensional spatiotemporal attack network. The global distribution of attack event timestamps refers to the changes in attack rhythm (e.g., a sudden shift from reconnaissance scanning to vulnerability exploitation) identified by the large language model through temporal reasoning, providing a basis for subsequent judgment of attack-defense escalation points. The division of roles among different groups refers to the division of labor among reconnaissance personnel, exploiters, and lateral movement operators. Spatiotemporal reasoning is then performed on the global distribution of attack event timestamps and the division of roles among different groups to obtain the network's dynamic attack narrative characteristics. Furthermore, based on the dynamic attack narrative characteristics and the network's asset spatial layout (e.g., the node positions of each asset in the network) and the spatiotemporal characteristics of exploitable vulnerabilities (e.g., the time interval during which vulnerabilities are exploitable and their location in the network space), the target objects and attack types to be attacked in the future are predicted; the target objects refer to the assets to be attacked in the future. Finally, based on the spatiotemporal characteristics of all future target objects and their attack types, a comprehensive attack and defense representation report of the network is generated. The aforementioned attack protection panoramic characterization report may include, but is not limited to, the identities, roles, and active periods of attack gang members, a complete attack chain mapping table, asset risk levels, a visualized campaign topology map (such as a support force-oriented map and spatiotemporal overlay display), and a complete campaign timeline (such as event sequences and state transitions accurate to the second), thereby supporting the formulation of network front-end deployment strategies.
[0034] Please see Figure 2 As shown in this embodiment, a network attack attribution analysis system integrating knowledge graphs and large-scale models is provided. This network attack attribution analysis system integrating knowledge graphs and large-scale models includes: The path deconstruction module is used to acquire multimodal network information, deconstruct the path of the multimodal network information, and obtain the complete access trajectory. The access behavior determination module is used to determine rare access behaviors of the network based on the complete access trajectory. The access vector generation module is used to identify the access status of rare access behaviors from multiple attack sources and generate attack source access vectors. The gang tracing module is used to trace the attack gangs in the network based on cluster analysis of the access vectors of the attack sources. The attack channel identification module is used to construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph, and identify potential attack penetration channels. The attack network building module is used to construct a three-dimensional spatiotemporal attack network based on potential attack penetration channels; The attack narrative determination module is used to perform large language model analysis on three-dimensional spatiotemporal attack networks to obtain the dynamic attack narrative characteristics of the network. The attack prediction module is used to predict the future attack characteristics of a network based on dynamic attack narrative characteristics.
[0035] This network attack tracing and analysis system, which integrates knowledge graphs and large models, accurately reconstructs the complete attack chain and distinguishes the roles and responsibilities of different groups when facing gang-related attack scenarios. This is achieved through rare access behavior detection, knowledge graph association modeling, and large language model inference. As a result, it improves the security and reliability of network operations.
[0036] In another embodiment, the path deconstruction module is used to acquire multimodal network information, perform path deconstruction on the multimodal network information, and obtain a complete access trajectory, including: Based on the distribution of network activity time intervals, multimodal network information is collected and preprocessed; among which, multimodal network information includes Web access logs, API call records, and directory traversal behavior records; The multimodal network information is deconstructed to obtain several path elements, including protocol, domain name, interface level, operation type, and behavioral parameter characteristics. Based on these path elements, a complete access trajectory is constructed. The access behavior determination module is used to determine rare network access behaviors based on the complete access trajectory, including: Perform global frequency statistics on all complete access trajectories, remove normal business paths, and obtain several preliminary complete access trajectories. Inverse document frequency quantization is performed on several initially screened complete access trajectories to identify rare access behaviors on the network.
[0037] In another embodiment, the access vector generation module is used to identify the access status of rare access behaviors from multiple attack sources and generate attack source access vectors, including: Extract multidimensional temporal features of rare access behaviors; among which, multidimensional temporal features include access time interval, temporal variation pattern of behavioral parameters, and agent rotation time sequence; Based on multi-dimensional temporal features, the access behavior of multiple attack sources is matched to obtain multi-dimensional access attribute elements of multiple attack sources. Among them, the multi-dimensional access attribute elements include the path rarity sequence of rare access behaviors initiated by multiple attack sources, access request parameters, access time distribution, access session duration, and access location jump. Semantic vector transformation and dimensionality reduction are performed on multi-dimensional access attribute elements to obtain the attack source access vector; The gang tracing module is used to trace attack gangs in a network based on cluster analysis of attack source access vectors, including: A multi-attack source similarity matrix is constructed by performing composite metric processing based on the attack source access vector; the composite metric processing includes cosine similarity and Euclidean distance metrics. Based on the multi-attack source similarity matrix, the attack sources are clustered and merged to identify the core gang clusters and peripheral interference sources; the core gang clusters are corrected based on the peripheral interference sources to determine the network's attack gang groups.
[0038] In another embodiment, the attack channel determination module is used to construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph, and determine potential attack penetration channels, including: Obtain the multi-dimensional entity attributes corresponding to the attack group; among which, the multi-dimensional entity attributes include attack source identity attributes, target asset attributes, vulnerability attributes, and attack time attributes; import the multi-dimensional entity attributes into the graph engine to perform entity relationship definition and construct a multi-dimensional knowledge graph of the attack group; A path retrieval from the attack source to the target asset is performed on the multidimensional knowledge graph to obtain several feasible attack penetration channels; vulnerability remediation prediction is performed on each feasible attack penetration channel to determine whether each feasible attack penetration channel belongs to a potential attack penetration channel. The attack network building module is used to construct a three-dimensional spatiotemporal attack network based on potential attack penetration channels, including: By identifying shared resources and identifying the collaborative patterns of attack groups across all potential attack channels, a three-dimensional spatiotemporal attack network can be constructed.
[0039] In another embodiment, the attack narrative determination module is used to perform large language model analysis on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative features of the network, including: The global arrangement of attack event timestamps and the division of roles among gang members in a three-dimensional spatiotemporal attack network are extracted using a large language model. Spatiotemporal reasoning is performed on the global arrangement of attack event timestamps and the division of roles among gang members to obtain the dynamic attack narrative characteristics of the network. Among them, the dynamic attack narrative characteristics include the temporal and spatial collaborative relationships of multi-point attack behaviors launched by different gang members with different role labels. The attack prediction module is used to predict future attack characteristics of a network based on dynamic attack narrative features, including: Based on the dynamic attack narrative characteristics and the spatial layout of network assets and the spatiotemporal characteristics of exploitable vulnerabilities, predict the target objects of future attacks on the network and the types of attacks they will be subjected to. Based on the spatiotemporal characteristics of all future targets and the types of attacks they may be subjected to, a comprehensive attack and defense profile report of the network is generated.
[0040] The network attack tracing and analysis system that integrates knowledge graphs and large models of the present invention operates and has the same effect as the network attack tracing and analysis method that integrates knowledge graphs and large models described above. Therefore, the network attack tracing and analysis system that integrates knowledge graphs and large models will not be described again here.
[0041] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. This disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.
Claims
1. A network attack trace analysis method fusing a knowledge graph and a large model, characterized in that, include: Step S1: Obtain multimodal network information, perform path deconstruction on the multimodal network information, and obtain a complete access trajectory; Based on the complete access trajectory, rare access behaviors of the network are identified; Step S2: Identify the access status of the rare access behavior from multiple attack sources and generate an attack source access vector; based on the cluster analysis of the attack source access vector, trace the attack groups of the network. Step S3: Construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph to determine potential attack penetration channels; construct a three-dimensional spatiotemporal attack network based on the potential attack penetration channels; Step S4: Perform large language model analysis on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative features of the network; Based on the dynamic attack narrative characteristics, predict the future attack characteristics of the network.
2. The network attack attribution analysis method integrating knowledge graphs and large models as described in claim 1, characterized in that: In step S1, multimodal network information is acquired, and path deconstruction is performed on the multimodal network information to obtain a complete access trajectory; Based on the complete access trajectory, rare network access behaviors are identified, including: Based on the distribution of network activity time intervals, multimodal network information is collected and preprocessed; wherein, the multimodal network information includes Web access logs, API call records, and directory traversal behavior records; The multimodal network information is deconstructed to obtain several path elements; wherein, the several path elements include protocol, domain name, interface level, operation type and behavioral parameter characteristics; based on the several path elements, a complete access trajectory is constructed; Perform global frequency statistics on all complete access trajectories, remove normal business paths, and obtain several preliminary complete access trajectories. Inverse document frequency quantization is performed on the several initial screening complete access trajectories to identify rare access behaviors of the network.
3. The network attack attribution analysis method integrating knowledge graphs and large models as described in claim 1, characterized in that: In step S2, the rare access behavior is identified by multiple attack source access status, and an attack source access vector is generated. Based on cluster analysis of the attack source access vectors, the attack groups of the network are traced, including: Extract multidimensional temporal features of the rare access behaviors; wherein, the multidimensional temporal features include access time interval, temporal variation law of behavioral parameters, and agent rotation time sequence; Based on the multi-dimensional temporal features, the access behavior of multiple attack sources is matched to obtain multi-dimensional access attribute elements of multiple attack sources; wherein, the multi-dimensional access attribute elements include the path rarity sequence of rare access behaviors initiated by multiple attack sources, access request parameters, access time distribution, access session duration, and access location jump. The multi-dimensional access attribute elements are subjected to semantic vector transformation and dimensionality reduction to obtain the attack source access vector. A multi-attack source similarity matrix is constructed by performing composite metric processing on the attack source access vector; wherein, the composite metric processing includes cosine similarity and Euclidean distance metrics; Based on the multi-attack source similarity matrix, the attack sources are clustered and merged to identify core gang clusters and peripheral interference sources; the core gang clusters are corrected based on the peripheral interference sources to determine the attack gang groups of the network.
4. The network attack attribution analysis method integrating knowledge graphs and large models as described in claim 1, characterized in that: In step S3, a multidimensional knowledge graph of the attack group is constructed, and path retrieval is performed on the multidimensional knowledge graph to determine potential attack penetration channels; Based on the potential attack penetration channels, a three-dimensional spatiotemporal attack network is constructed, including: Obtain the multi-dimensional entity attributes corresponding to the attack group; wherein, the multi-dimensional entity attributes include attack source identity attributes, target asset attributes, vulnerability attributes, and attack time attributes; import the multi-dimensional entity attributes into a graph engine to perform entity relationship definition and construct a multi-dimensional knowledge graph of the attack group; A path retrieval from the attack source to the target asset is performed on the multidimensional knowledge graph to obtain several feasible attack penetration channels; vulnerability remediation prediction is performed on each feasible attack penetration channel to determine whether each feasible attack penetration channel belongs to a potential attack penetration channel. By identifying shared resources and identifying the collaborative patterns of the attack groups among all potential attack channels, a three-dimensional spatiotemporal attack network can be constructed.
5. The network attack attribution analysis method integrating knowledge graphs and large models as described in claim 1, characterized in that: In step S4, a large language model analysis is performed on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative features of the network; Based on the dynamic attack narrative characteristics, predict the future attack characteristics of the network, including: The global arrangement of attack event timestamps and the division of roles among gang members in the three-dimensional spatiotemporal attack network are extracted using a large language model. Spatiotemporal reasoning is performed on the global arrangement of the attack event timestamps and the division of roles among the gang members to obtain the dynamic attack narrative features of the network; wherein, the dynamic attack narrative features include the temporal and spatial collaborative relationships of multi-point attack behaviors initiated by different gang members with different role labels. Based on the dynamic attack narrative characteristics and the spatiotemporal characteristics of the network's asset spatial layout and exploitable vulnerabilities, predict the target objects that the network will be attacked in the future and the types of attacks they will be subjected to. Based on the spatiotemporal characteristics of all future targets and the types of attacks they are subjected to, a comprehensive attack protection profile report for the network is generated.
6. The network attack trace analysis system fusing knowledge graph and large model, characterized in that, include: The path deconstruction module is used to acquire multimodal network information, deconstruct the path of the multimodal network information, and obtain a complete access trajectory. The access behavior determination module is used to determine rare access behaviors of the network based on the complete access trajectory. The access vector generation module is used to identify the access status of the rare access behavior from multiple attack sources and generate an attack source access vector. The gang tracing module is used to trace the attack gangs of the network based on the clustering analysis of the access vectors of the attack sources; The attack channel determination module is used to construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph, and determine potential attack penetration channels. An attack network construction module is used to construct a three-dimensional spatiotemporal attack network based on the potential attack penetration channels. The attack narrative determination module is used to perform large language model analysis on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative characteristics of the network. An attack prediction module is used to predict the future attack characteristics of the network based on the dynamic attack narrative characteristics.
7. The network attack attribution analysis system integrating knowledge graphs and large models as described in claim 6, characterized in that: The path deconstruction module is used to acquire multimodal network information, perform path deconstruction on the multimodal network information, and obtain a complete access trajectory, including: Based on the distribution of network activity time intervals, multimodal network information is collected and preprocessed; wherein, the multimodal network information includes Web access logs, API call records, and directory traversal behavior records; The multimodal network information is deconstructed to obtain several path elements; wherein, the several path elements include protocol, domain name, interface level, operation type and behavioral parameter characteristics; based on the several path elements, a complete access trajectory is constructed; The access behavior determination module is used to determine rare access behaviors of the network based on the complete access trajectory, including: Perform global frequency statistics on all complete access trajectories, remove normal business paths, and obtain several preliminary complete access trajectories. Inverse document frequency quantization is performed on the several initial screening complete access trajectories to identify rare access behaviors of the network.
8. The network attack attribution analysis system integrating knowledge graphs and large models as described in claim 6, characterized in that: The access vector generation module is used to identify the access status of the rare access behavior from multiple attack sources and generate an attack source access vector, including: Extract multidimensional temporal features of the rare access behaviors; wherein, the multidimensional temporal features include access time interval, temporal variation law of behavioral parameters, and agent rotation time sequence; Based on the multi-dimensional temporal features, the access behavior of multiple attack sources is matched to obtain multi-dimensional access attribute elements of multiple attack sources; wherein, the multi-dimensional access attribute elements include the path rarity sequence of rare access behaviors initiated by multiple attack sources, access request parameters, access time distribution, access session duration, and access location jump. The multi-dimensional access attribute elements are subjected to semantic vector transformation and dimensionality reduction to obtain the attack source access vector. The gang tracing module is used to trace the attack gangs of the network based on cluster analysis of the attack source access vectors, including: A multi-attack source similarity matrix is constructed by performing composite metric processing on the attack source access vector; wherein, the composite metric processing includes cosine similarity and Euclidean distance metrics; Based on the multi-attack source similarity matrix, the attack sources are clustered and merged to identify core gang clusters and peripheral interference sources; the core gang clusters are corrected based on the peripheral interference sources to determine the attack gang groups of the network.
9. The network attack attribution analysis system integrating knowledge graphs and large models as described in claim 6, characterized in that: The attack channel determination module is used to construct a multi-dimensional knowledge graph of the attack group, perform path retrieval on the multi-dimensional knowledge graph, and determine potential attack penetration channels, including: Obtain the multi-dimensional entity attributes corresponding to the attack group; wherein, the multi-dimensional entity attributes include attack source identity attributes, target asset attributes, vulnerability attributes, and attack time attributes; import the multi-dimensional entity attributes into a graph engine to perform entity relationship definition and construct a multi-dimensional knowledge graph of the attack group; A path retrieval from the attack source to the target asset is performed on the multidimensional knowledge graph to obtain several feasible attack penetration channels; vulnerability remediation prediction is performed on each feasible attack penetration channel to determine whether each feasible attack penetration channel belongs to a potential attack penetration channel. The attack network construction module is used to construct a three-dimensional spatiotemporal attack network based on the potential attack penetration channels, including: By identifying shared resources and identifying the collaborative patterns of the attack groups among all potential attack channels, a three-dimensional spatiotemporal attack network can be constructed.
10. The network attack attribution analysis system integrating knowledge graphs and large models as described in claim 6, characterized in that: The attack narrative determination module is used to perform large language model analysis on the three-dimensional spatiotemporal attack network to obtain the dynamic attack narrative features of the network, including: The global arrangement of attack event timestamps and the division of roles among gang members in the three-dimensional spatiotemporal attack network are extracted using a large language model. Spatiotemporal reasoning is performed on the global arrangement of the attack event timestamps and the division of roles among the gang members to obtain the dynamic attack narrative features of the network; wherein, the dynamic attack narrative features include the temporal and spatial collaborative relationships of multi-point attack behaviors initiated by different gang members with different role labels. The attack prediction module is used to predict the future attack characteristics of the network based on the dynamic attack narrative characteristics, including: Based on the dynamic attack narrative characteristics and the spatiotemporal characteristics of the network's asset spatial layout and exploitable vulnerabilities, predict the target objects that the network will be attacked in the future and the types of attacks they will be subjected to. Based on the spatiotemporal characteristics of all future targets and the types of attacks they are subjected to, a comprehensive attack protection profile report for the network is generated.
