A Threat Hunting and Origin Tracing System and Method Based on Neural Symbolic Architecture

By using a threat hunting and tracing system based on a neural symbolic architecture, and leveraging a structured search space and attack pattern knowledge base, combined with heuristic search algorithms and physical matching verification, a trustworthy attack chain is generated. This solves the problems of attack chain completion and the credibility of tracing results in anti-forensic scenarios, achieving highly robust threat hunting and tracing.

CN123122931APending Publication Date: 2026-11-17HARBIN INST OF TECH AT WEIHAI +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611339087.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot simultaneously ensure the attack chain completion capability and the high reliability and auditability of the tracing results in anti-forensic scenarios, making it difficult to meet the needs of automated, high-precision, and highly robust threat hunting and tracing.

Method used

A threat hunting and tracing system based on a neural symbolic architecture is adopted. Through a structured search space, an attack pattern knowledge base, and an attack chain search module, a heuristic search algorithm and physical matching verification are used to generate a trusted attack chain, ensuring that the reasoning process is bounded and controllable, and meeting the compliance requirements of network security auditing and forensics.

Benefits of technology

In anti-forensic scenarios, it significantly improves the robustness of tracing, eliminates the fabrication of causal relationships for attacks, ensures the controllability of the reasoning process, and provides a complete traceability and compliant acceptance of the output evidence chain, thus solving the problem of broken tracing links.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122931A_ABST
    Figure CN123122931A_ABST
Patent Text Reader

Abstract

This application provides a threat hunting and attribution system and method based on a neural symbolic architecture. The system includes a structured search space for storing observation attribution graphs that conform to temporal causal constraints; an attack pattern knowledge base for storing attack pattern graphs; and an attack link search module that, in response to threat hunting and attribution requests, searches for complete attack links from the starting anchor point to the ending anchor point by executing a heuristic search algorithm based on a neural symbolic architecture. In each link expansion operation of the algorithm, virtual attack sub-chains are inferred under the constraints of prior attack patterns, and the semantic attributes of their virtual target nodes are filled only with the inference results. Trustworthy attack sub-chains are selected based on the physical matching verification between the multi-constraint search results and the virtual sub-chains. When no trustworthy attack sub-chains can be selected, the virtual attack sub-chains are used as inferred attack sub-chains and their bias cost is evaluated.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of network security technology, specifically, it provides a threat hunting and tracing system and method based on neural symbolic architecture. Background Technology

[0002] With the digital transformation of enterprises, the audit logs and security alerts handled by security operations centers are characterized by massive, heterogeneous, multi-source, fragmented, and highly redundant data. The network attack methods they face are becoming increasingly complex, covert, and persistent, with sophisticated network attacks, such as advanced persistent threats (APS), becoming a major threat to enterprise cybersecurity. In existing security operations systems, pre-emptive security technologies such as alert analysis and situational awareness are relatively mature, effectively filtering large amounts of redundant alerts, alleviating alert fatigue, and accurately identifying high-priority suspicious security alert logs, thus providing basic security capabilities for attack existence determination. However, in practical applications, pre-emptive alert screening capabilities can only identify risks and cannot reconstruct the complete attack chain. The key bottleneck in security operations has shifted from "alert identification" to the "threat hunting and attribution" stage.

[0003] Threat hunting and attribution refers to security operation techniques that proactively discover potential attack threats and reconstruct the complete attack chain by analyzing multi-source heterogeneous security monitoring data from enterprises, thus restoring the entire attack process. Currently, the mainstream key technologies for threat hunting and attribution fall into two categories: rule matching technology based on deterministic attribution graphs and threat hunting and attribution technology based on Large Language Models (LLMs). These two attribution technologies have significant technical shortcomings and are not compatible or complementary: deterministic attribution graph rule matching technology is reliable when logs are complete, but it is extremely vulnerable and fails to attribution in anti-forensics or log-missing scenarios; unconstrained LLMs attribution technology can fill in missing links, but the reasoning process is uncontrollable, the results are easily fabricated, and there is no audit credibility.

[0004] Therefore, existing technologies generally cannot simultaneously ensure the attack chain completion capability and the high reliability and auditability of the tracing results in anti-forensic scenarios, making it difficult to meet the needs of automated, high-precision, and highly robust threat hunting and tracing. Summary of the Invention

[0005] This application provides a threat hunting and tracing system based on a neural symbolic architecture through embodiments, including: A structured search space is used to store observation source maps that conform to time-series causal constraints, generated based on raw monitoring data; An attack pattern knowledge base is used to store attack pattern graphs. The attack pattern graphs are consistent with the observation and tracing graphs in terms of node types and edge semantics, including attack primitives that represent various standardized attack patterns. The attack chain search module, in response to threat hunting and attribution requests, searches for complete attack chains from the starting anchor point to the ending anchor point by executing a heuristic search algorithm based on neural symbolic architecture. In each chain expansion operation of the algorithm, a new trusted attack sub-chain or inferred attack sub-chain is generated starting from the current node of the current candidate attack chain based on the following steps: Using attack primitives that exist in the attack pattern knowledge base and conform to the environmental characteristics of the current node as inference templates, virtual target nodes are inferred. Virtual target nodes are populated using only the semantic attribute information of nodes obtained through inference to generate virtual attack subchains; A multi-constraint search is performed on the structured search space, and the search results are physically matched and verified with the virtual subchains to filter out the trustworthy attack subchains. When a trustworthy attack subchain cannot be identified, the virtual attack subchain is used as the inference attack subchain, and its bias cost is evaluated.

[0006] This application also provides a threat hunting and attribution method based on a neural symbolic architecture through embodiments. The method uses the aforementioned system for threat hunting and attribution, including: In response to threat hunting and attribution requests, attribution boundaries are determined, including attribution start and end anchor points; A heuristic attack chain search algorithm based on neural symbolic architecture is executed to search for each complete attack chain from the starting anchor point to the ending anchor point. The complete attack chain includes each attack sub-chain connected sequentially between the starting anchor point and the ending anchor point.

[0007] This application provides a threat hunting and attribution method and system based on neural symbolic architecture through its embodiments. These improvements address the core shortcomings of existing technologies, such as attribution failure, uncontrollable reasoning, and unreliable results in anti-forensic adversarial scenarios, thereby achieving the following: (1) By constraining the model reasoning with prior attack knowledge, the illusion problem of large language model is suppressed from the mechanism level, the fabrication of attack causal relationship is eliminated, and the reasoning process is bounded and controllable; (2) A neural symbolic architecture that decouples semantic generation and physical verification of inference model is adopted to perform deterministic physical matching verification on all candidate links generated by inference. Based on the matching results, the physical verification credible links and inference clue links are strictly distinguished. The output evidence links have complete traceability and compliance acceptance capabilities, which meet the compliance requirements of network security audit and evidence collection.

[0008] (3) The problem of tracing the source under the conditions of incomplete logs and lack of evidence is modeled as a partially observable cost-constrained heuristic graph search problem, which breaks through the technical bottleneck of the traditional deterministic tracing graph when the audit log is incomplete and the tracing link is broken and the process is interrupted, and significantly improves the robustness of tracing in complex and highly adversarial scenarios. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the architecture of a threat hunting and tracing system based on a neural symbolic architecture provided in the embodiments of this application; Figure 2 This is a flowchart illustrating the threat hunting and tracing method based on neural symbolic architecture provided in the embodiments of this application; Figure 3 This is a schematic diagram of attack primitives stored in the attack pattern knowledge base provided according to the embodiments of this application; Figure 4 This is a schematic diagram of the architecture of a threat hunting and tracing system based on a neural symbolic architecture provided in the embodiments of this application; Figure 5 This is a flowchart illustrating the threat hunting and tracing method based on neural symbolic architecture provided in the embodiments of this application; Figure 6 This is a schematic diagram of the architecture of a threat hunting and tracing system based on a neural symbolic architecture provided in the embodiments of this application; Figure 7 This is a schematic diagram of a virtual attack subchain provided according to an embodiment of this application; Figure 8 This is a schematic diagram of alternative attack subchains provided according to embodiments of this application; Figure 9 This is a schematic diagram of a trusted attack subchain provided according to an embodiment of this application; Figure 10 This is a schematic diagram illustrating link extension via a trusted attack subchain according to an embodiment of this application; Figure 11 This is a schematic diagram of an inference attack subchain provided according to an embodiment of this application; Figure 12 This is a schematic diagram illustrating link extension via inference attack subchains according to an embodiment of this application; Figure 13 This is a schematic diagram of the architecture of a threat hunting and tracing system based on a neural symbolic architecture provided in the embodiments of this application; Figure 14 This is a flowchart of a heuristic path search algorithm based on a neural symbolic architecture provided according to an embodiment of this application. Detailed Implementation

[0010] The present application will now be further described based on preferred embodiments and with reference to the accompanying drawings.

[0011] As described in the background section, current mainstream threat hunting and attribution methods have limitations when used in anti-forensic scenarios.

[0012] Specifically, when rule-matching technology based on deterministic source graphs faces anti-forensic techniques such as log erasure, log tampering, fileless execution, memory residency, and segmented lurking, the complete record of system calls is directly destroyed, resulting in problems such as breaks, fragmentation, and node disconnection in the deterministic source graph. The complete attack chain is fragmented into multiple unconnected local subgraphs. At this time, the deterministic source graph algorithm will directly terminate the source tracing process due to the link interruption, and it will be unable to deduce the key links missing in the logs. It is necessary to rely on analysts to manually fill in the link based on historical threat intelligence. Therefore, its automated source tracing capability is completely ineffective and cannot be adapted to highly adversarial anti-forensic scenarios.

[0013] To address issues such as missing logs and broken links in anti-forensic scenarios, the industry has proposed threat hunting and attribution technology based on Large Language Models (LLM). This technology relies on the zero-sample semantic understanding and generation capabilities of LLM to break through the fixed paradigm of traditional rule matching. It can autonomously reason and complete potential attack causal relationships based on fragmented alarm traces and local log information. However, this technology has an unavoidable credibility defect: LLM equates threat attribution tasks with open text generation tasks, without introducing physical constraints at the operating system level, formal verification boundaries, and temporal causal rules. It lacks system-level time awareness and logical verification mechanisms. Therefore, the model is prone to serious cognitive uncertainty during the reasoning process. In order to ensure the coherence of the attack narrative, it may develop excessive illusions, autonomously fabricate non-existent system calls, causal relationships, and attack behaviors, and generate a large number of false attack links.

[0014] At the same time, this type of technology lacks a standardized verification mechanism, and the output results cannot be verified for authenticity. It does not meet the requirements of network security compliance audits for the authenticity, objectivity, and admissibility of evidence, and cannot be applied to formal security assessment, attack forensics, and source tracing work. Its practicality and reliability are seriously insufficient.

[0015] To address the problems inherent in the two types of threat hunting and attribution technologies mentioned above in scenarios involving anti-forensics and fragmented attribution maps, this application provides a threat hunting and attribution system based on a neural symbolic architecture through embodiments.

[0016] like Figure 1As shown, the threat hunting and tracing system based on neural symbolic architecture provided in this application includes at least a structured search space, an attack pattern knowledge base, and an attack link search module. The structured search space and the attack pattern knowledge base constitute the data layer of the threat hunting and tracing system, while the attack link search module serves as the core functional module for threat hunting and tracing, performing specific threat hunting and tracing operations.

[0017] The structured search space is used to store directed acyclic graphs that conform to temporal causal constraints, generated based on raw monitoring data (such as enterprise multi-source heterogeneous monitoring logs). In this application, this temporally ordered and structurally standardized graph space built on monitoring data is called the source graph. The source graph constitutes the data foundation for the attack chain search module described later to search for each node and attack edge. It is understandable that in anti-forensic scenarios, the raw monitoring data may be incomplete due to the attacker's anti-forensic tampering, log clearing, behavior concealment, and other adversarial measures. Therefore, the source graph generated based on the raw monitoring data is called the observation source graph or the actual source graph, which can be considered a subgraph of the ideal source graph. The attack pattern knowledge base is used to store attack pattern graphs. The attack pattern graphs are consistent with the source graphs in terms of node types and edge semantics. They are composed of attack primitives that represent various standardized attack patterns. In the attack link search process described later, the attack patterns searched from the attack pattern knowledge base can provide a priori attack paradigm templates and constraints for the reasoning process from the current node to the next node, so as to avoid the reasoning model from generating reasoning illusions and fabricating attack causal relationships.

[0018] It is understandable that searching a structured search space essentially means searching the source graphs stored within it; similarly, searching an attack pattern knowledge base essentially means searching the attack pattern graphs stored within it. Therefore, in the embodiments of this application, source graphs and structured search spaces, as well as their physical storage entities, such as local databases or cloud databases storing source graphs, can be considered as different implementations of the same concept; and attack pattern graphs and attack pattern knowledge bases, as well as their physical storage entities, such as local databases or cloud databases storing attack pattern graphs, can be considered as different implementations of the same concept.

[0019] The attack chain search module is the core of this threat hunting and attribution system. In response to threat hunting and attribution requests, this module searches for the complete attack chain from the starting anchor point to the ending anchor point by executing a heuristic search algorithm based on neural symbolic architecture.

[0020] This application also provides a threat hunting and attribution method based on neural symbolic architecture, which uses, for example... Figure 1 The system shown is used for threat hunting and tracing. Figure 2A flowchart of the method according to some embodiments is shown, with reference to Figure 2 The method includes the following steps: Step 100: In response to a threat hunting and tracing request, determine the tracing boundary, which includes a tracing start anchor point and a tracing end anchor point; Step 200: Execute a heuristic attack chain search algorithm based on neural symbolic architecture to search for each complete attack chain from the starting anchor point to the ending anchor point. The complete attack chain includes each attack sub-chain connected sequentially between the starting anchor point and the ending anchor point.

[0021] The following detailed description of the threat hunting and tracing system and method based on neural symbolic architecture provided in this application, in conjunction with the accompanying drawings and specific embodiments, is provided in detail.

[0022] <Establishment and management of structured search space and determination of threat attribution boundaries> Enterprises need to monitor various aspects of their daily operations, generating multi-source heterogeneous monitoring data, such as endpoint detection and response kernel event streams, operating system event trace (ETW) data, network traffic metadata, and virtualization monitoring trajectories. Generally, this raw, multi-source heterogeneous monitoring data is stored in the form of unstructured monitoring logs. In order to support the systematic search and reasoning of attack chains during threat hunting and attribution, it is necessary to convert these raw data in different formats into a standardized graph model that conforms to the laws of time causality, i.e., a structured search space, so as to provide a unified, physically interpretable structured data foundation for threat hunting and attribution.

[0023] In some embodiments, such as Figure 1 As shown, the system also includes a search space management module for establishing and managing the structured search space.

[0024] Specifically, the search space management module continuously collects multi-source heterogeneous monitoring data from the enterprise environment, including endpoint detection and response (EDR) kernel event streams, operating system event trace (ETW) data, network traffic metadata, and virtualization monitoring trajectories. Based on the collected multi-source heterogeneous monitoring data, it follows temporal causal constraints. Perform loop removal processing to eliminate structures such as process loops and handle reuse, and transform the monitoring data into a directed acyclic graph under causal constraints. This forms a structured search space, in which... for The set of nodes, for A set of directed edges that follow temporal causal constraints.

[0025] It should be noted that in anti-forensic scenarios, some logs may be missing due to attackers performing anti-forensic operations. It is a complete theoretical origin diagram A subset of, i.e. Both share the same node and edge type system, but There may be missing nodes or edges due to anti-forensic operations performed by the attacker. Therefore, in the embodiments of this application, nodes or edges may also be missing. This is called the observational source map or the actual source map.

[0026] In some embodiments, the search space management module can predefine the system ontology according to the operating system's operating specifications. ,Establish The document defines the types of nodes in the central graph and defines a physical identifier (ID) field (with unique values) and a semantic attribute field (attr) for each type of node, which are used to materialize various processes, files, and other records in the monitoring data. The unified rules adopted when dealing with nodes in the system.

[0027] For example, graph nodes can be divided into process nodes. File nodes Network nodes and registry nodes The physical identifiers and semantic attribute definitions of the four types of nodes are shown in Table 1: Table 1 Similarly, in some embodiments, the search space management module can also map various discrete operational behaviors in the system into a standardized set of operating system API actions. To materialize the operational behaviors in the monitoring data into The unified rules for directed edges in the context of the problem.

[0028] For example, it can be In this context, directed edges are defined as causal transformation structures: , in , for timestamp, Represents a node At any moment Through standardized actions Acting on nodes In this application, without loss of generality, causal edges are referred to as... The node that initiates the action The object toward which the action is directed is called the source node. Called the target node, therefore, the observation source map This refers to a causal graph with a band type that conforms to the system ontology specification, obtained by performing loop removal processing on multi-source heterogeneous monitoring data in the enterprise environment in accordance with temporal causal constraints.

[0029] Clearly, as the enterprise operates, the content of multi-source heterogeneous monitoring logs is continuously added. Correspondingly, the search space management module needs to perform causal constraint checks on the newly added log content in real-time, periodically, or when trigger conditions are met (such as the added volume exceeding a preset threshold). The additional operations are used to continuously maintain and update the structured search space.

[0030] In some embodiments, in addition to managing the structured search space, such as Figure 2 As shown, the search space management module can also serve as the entry point for threat hunting and tracing of the entire system. In step 100, in response to threat hunting and tracing requests (automatically triggered or manually submitted), the boundaries of threat hunting and tracing are determined.

[0031] For example, when the system is attacked, the security operations center triggers a security alert log. This security alert log is received by the search space management module, which parses it and instantiates the compromised entity indicated by the alert as the starting anchor point of the attack chain (i.e., the threat). Determine the termination anchor point of the attack chain by combining the investigation context. This involves defining the boundaries for threat hunting and attribution, thereby defining the mission objectives of threat hunting and attribution as follows: In the observation of the source map In the process, using the starting and ending anchor points as boundaries, a complete attack chain that satisfies the physical laws of the operating system is reconstructed. ,in, As an intermediate node, it is obvious that... , ,…, Both have causal edges The isomorphic form, which includes source node, action, and target node information, represents an attack action and the two nodes connected to it in a complete attack chain. Therefore, in the embodiments of this application, the graphical structure composed of source node, attack action, and target node in the above-mentioned attack chain is called an attack sub-chain. Obviously, a complete attack chain... It consists of several attack subchains that satisfy causal constraints connected sequentially.

[0032] In addition to proactively triggering threat hunting and attribution, the search space management module can also respond to manually initiated threat hunting and attribution requests, determining attribution boundaries. For example, users can input or upload information about discovered traps through a graphical user interface, and the search space management module will determine the starting anchor point of the attack chain based on the user's input or uploaded information. Determine the termination anchor point of the attack chain by combining the investigation context. This allows us to define the boundaries of threat hunting and attribution in response to user requests.

[0033] Taking an APT intrusion incident as an example: The security operations center received an alert indicating that the process powershell.exe (PID=2384, startup time=14:02:31, HostID=H01) on host-01 had suspicious outbound behavior, and the system instantiated it as the starting anchor. Based on the investigation context, the critical configuration file \\Windows\\System32\\config\\SAM on the attacker's lateral movement target host Host-02 was identified as the termination anchor. Because the attacker cleared part of the kernel event log, the source tracing diagram was observed. There is a break between the file written by powershell.exe and the access to the SAM file. It is necessary to fill in the missing intermediate nodes based on the search results of the complete attack chain by the attack chain search module.

[0034] <Establishment and Management of Attack Pattern Knowledge Base> In some embodiments, such as Figure 1 As shown, the system also includes a knowledge base management module, used to establish and manage an attack pattern knowledge base that stores attack pattern graphs. Specifically, the knowledge base management module can continuously collect unstructured threat intelligence data such as the MITRE ATT&CK tactical framework, Atomic Red Team attack scripts, and publicly available APT incident reports. It extracts structured attack primitives from the collected threat intelligence data and aggregates them into attack pattern graphs, thereby establishing an attack pattern knowledge base.

[0035] In this application, the attack pattern diagram This refers to a typed causal graph extracted from multi-source threat intelligence data that conforms to the system ontology specification. This refers to the set of nodes involved in various extracted attack actions. For each type of extracted attack action, there is a causal edge. Each causal edge represents an attack action that points from the source node to the target node. Therefore, the attack pattern graph and the source graph are consistent in terms of node type and causal edge semantics.

[0036] Specifically, a large language model can be used with fixed JSON output constraints to convert unstructured threat intelligence into different attack primitives. Primitives that do not conform to the ontology specification can then be filtered out, resulting in typed attack primitives that conform to the system ontology specification. Each attack primitive represents a known and system-compatible attack pattern, which can be represented as follows: Figure 3 The single-jump edge action triple shown is (source node, action, target node).

[0037] Furthermore, the knowledge base management module can vectorize the attack primitives extracted from unstructured threat intelligence. Specifically, each attack primitive is serialized into canonical text, and then the serialized text is encoded into high-dimensional feature vectors (e.g., 3072-dimensional) using models such as text-embedding-3-large, thereby establishing a mapping relationship between the structured graph form and the feature vector form of each attack primitive. The generated high-dimensional feature vectors of each attack primitive can be stored in a vector database (such as FAISS) to complete the construction of the attack pattern knowledge base. By transforming unstructured threat intelligence into attack pattern graphs compatible with attribution graphs and storing the generated attack pattern graphs in a vectorized manner, searchable attack prior knowledge can be provided to support the inference of attack subchains during the attack chain search process described later.

[0038] Understandably, after building the attack pattern knowledge base, the knowledge base management module can also periodically or irregularly acquire new threat intelligence data to achieve continuous updates to the attack pattern knowledge base.

[0039] <Attack Link Search Module and Neural Symbolic Architecture> like Figure 1 and Figure 2 As shown, the attack link search module is the core functional module of this threat hunting and tracing system. This module is based on the observation and tracing map. (Structured search space) and attack pattern diagram Supported by the data layer comprised of the (attack pattern knowledge base), in step 200, a heuristic search algorithm based on neural symbolic architecture is executed at a given starting anchor point. and termination anchor point Under the condition that the search starts from arrive The complete attack chain. It is understood that the search results can be 0, 1, or multiple complete attack chains. Each complete attack chain obtained from a search, as described above, includes sequential connections to... and The various attack subchains between , ,…, .

[0040] Heuristic pathfinding algorithms are algorithms that find the optimal path from a known starting point to a known ending point in a graph structure. The basic idea is to start from the starting point and repeatedly perform new node generation, verification and path expansion operations until the expanded path reaches the ending point. In order to reduce the computational scale, after each loop, path filtering, pruning and other operations can be performed according to the preset bundle width to control the number of paths involved in each round of expansion.

[0041] Reconstructing the complete attack chain using a heuristic pathfinding algorithm starts from the initial anchor point. Starting from this point, the attack subchain is continuously expanded. The expansion method is as follows: for each subchain that has not yet reached the termination anchor point... The candidate attack chain is used as the source node, with its current node (i.e., the target node of the newly added attack sub-chain after the most recent expansion) as the source node. The search for reachable attack edges and target nodes is performed to expand the attack sub-chain. This process is repeated until the path search is successful, i.e., the termination anchor point is reached. Or it may trigger a failure condition, such as the candidate attack link reaching the termination anchor point during the search process. Previously, the cost of the deviation had exceeded an acceptable range, leading to the abandonment of further expansion of the candidate attack link.

[0042] In anti-forensic adversarial scenarios, due to the actual observation and tracing of the source map The data may be fragmented and lack information. Therefore, it is necessary to use inference models to infer attack subchains to complete the attack chain. However, as mentioned in the background, relying solely on neural network inference models (such as large language models) to expand attack subchains has inherent defects: when inferring new target nodes and edges from source nodes to target nodes (corresponding to attack actions), inference models such as large language models only perform probability sampling based on linguistic statistical rules and cannot perceive the physical constraints of the operating system. This may generate attack actions and targets that seem reasonable in semantics but are impossible at the physical level. For example, generating non-existent file path formats, process nodes that violate the rules of parent-child process creation, or causal edges with timestamps earlier than their parent nodes. In addition, the pure inference method lacks an active verification mechanism for actual monitoring data and cannot distinguish between "confirmed chains supported by existing physical evidence" and "hypothetical chains inferred solely by the model". This results in the inability to quantify the credibility of the final output attack chain and the risk of confusing inference illusions with objective facts.

[0043] To address the aforementioned issues, in the embodiments of this application, a neural symbol mechanism is introduced into the search algorithm executed by the attack chain search module to ensure that the generated complete attack chain conforms to the physical laws of the operating system. Accordingly, this new search algorithm is called a heuristic path search algorithm based on neural symbol architecture.

[0044] Specifically, the neural symbolic architecture refers to generating a new trusted attack subchain or inference attack subchain starting from the current node of the current candidate attack chain in each link expansion operation of the algorithm based on the following steps: Using attack primitives that exist in the attack pattern knowledge base and conform to the environmental characteristics of the current node as inference templates, virtual target nodes are inferred. Virtual target nodes are populated using only the semantic attribute information of nodes obtained through inference to generate virtual attack subchains; A multi-constraint search is performed on the structured search space, and the search results are physically matched and verified with the virtual subchains to filter out the trustworthy attack subchains. When a trustworthy attack subchain cannot be identified, the virtual attack subchain is used as the inference attack subchain, and its bias cost is evaluated.

[0045] In the above steps, firstly, attack primitives that conform to the current node's attack pattern are retrieved from the attack pattern graph. Symbolic constraints are applied to the output space of the inference model to prevent it from arbitrarily inferring attack actions and node types that deviate too much from the current attack phase pattern, thus compressing the range of erroneous assumptions from the source. Then, in the process of generating virtual attack subchains, only the semantic attribute information of nodes in the inference results is used to complete the information of virtual target nodes. No physical unique identifiers that can be placed on the system are generated out of thin air, which can effectively prevent the inference model from forging physical identifiers and causing error propagation. Furthermore, relying on observable system log information that is difficult for attackers to tamper with, candidate attack subchains are searched under multiple constraints. The search results and virtual attack subchains are rigidly verified using physical dependency matching rules that conform to causality to screen out credible attack subchains with real data support. Finally, when physical matching fails due to missing logs, data corruption, or other reasons, the inference results are used as inference attack subchains to expand the current candidate attack chain. The deviation of the chain after expansion by inference attack subchains is quantitatively evaluated according to the deviation cost evaluation rules to ensure that the further expansion of low-quality inference chains is stopped in time. The above steps work together to form a neural symbol-based link extension mechanism, enabling the reconstruction of attack links that combines semantic reasoning capabilities with physical credibility.

[0046] In some alternative embodiments, such as Figure 4As shown, the threat hunting and tracing system also includes three modules that can be called by the attack chain search module: a virtual attack subchain inference module, a trusted attack subchain matching module, and an inference attack subchain evaluation module. Figure 5 The diagram illustrates how the attack link search module expands the link by calling these three functional modules during step 200, which is the heuristic path search algorithm based on neural symbolic architecture. It can be seen that, supported by the data layer (structured search space, attack pattern knowledge base), each module of the functional layer performs its corresponding function, thereby realizing each step in the threat hunting and tracing method.

[0047] It is understood that, in the embodiments of this application, for Figure 1 , Figure 4 , Figure 5 The functional modules in the diagram should be interpreted broadly. They can be understood as traditional software modules composed of executable program code capable of performing corresponding functions, or as intelligent agents endowed with the skills required to autonomously perform corresponding functions, or other software or hardware implementations capable of achieving the corresponding functions. For example, Figure 6 The illustrated embodiment discloses a multi-agent architecture implementation of the threat hunting and tracing system. Agent 1 is the core agent, responsible for executing the aforementioned algorithm to search for attack chains. During the search process, agent 1 directs agents 4, 5, and 6 to perform virtual attack subchain inference, physical matching verification, screening of trusted attack subchains, and evaluation of inference subchain bias costs. Furthermore, agent 2 is responsible for managing and maintaining the structured search space and, upon receiving a threat hunting and tracing request, analyzes and determines the tracing boundary (starting anchor point). and termination anchor point The attack pattern knowledge base is then sent to agent1, while agent3 is responsible for managing and maintaining it. These agents are each given corresponding skills to cooperate and execute the attack pattern knowledge base. Figure 2 The threat hunting and tracing methods shown.

[0048] The following detailed description, in conjunction with specific embodiments, illustrates the detailed implementation of each module and step in the aforementioned neural symbolic architecture-based link extension mechanism.

[0049] a. Scenario-adaptive attack pattern retrieval and virtual attack subchain inference under pattern constraints During the execution of the attack link search algorithm, whenever it is necessary to perform a link expansion operation on the extracted current candidate attack link, the current node that exists in the attack pattern library and matches the candidate attack link is first selected. The attack primitives of the environment in which the node (i.e. the node farthest from the starting anchor point) is located serve as the inference template to infer the virtual target node.

[0050] Specifically, taking the current node of the current candidate attack link as an example. As the source node, extract its Skip Neighborhood ( Preferably equal to 2) and serialized into standard JSON format, mapped to a query vector using the same semantic encoding model, and then compared with the high-dimensional feature vectors corresponding to each attack primitive stored in the attack pattern knowledge base based on cosine similarity, retrieving the Top-K most similar ( Preferably, 8 primitives are used as the attack primitive set adapted to the current scenario. The method for calculating the cosine similarity between vectors is well known to those skilled in the art; for example, for any two vectors... and The cosine similarity is: , in, for and dot product, Let L be the modulus (L2 norm) of the vector.

[0051] Then, using the retrieved attack primitives (including source nodes, action edges, and target nodes) as structural templates, combined with the context, they are input into the inference model (such as LLM). Because the structural templates limit the attack actions and the types of target nodes (such as process nodes),... File nodes Network nodes and registry nodes Therefore, the inference model will generate target nodes under the constraints of prior attack patterns that conform to the current scenario, thereby effectively avoiding problems such as generating attack actions and target node types that do not conform to the current scenario or violate physical causality when the inference model performs unconstrained inference.

[0052] In the embodiments of this application, the inference model can be derived from the current node under the constraints of the structural template. The target node generated by the inference process is called a virtual target node. Regardless of whether the inference model infers the physical identifier ID information of the node, it only uses the semantic attribute information of the node obtained through inference to fill the virtual target node. The attr field, correspondingly, the directed graph structure obtained from the inference result above, consisting of "current node + action edge + virtual target node containing only node semantic attribute information", is called the virtual attack subchain. ,Right now .

[0053] By performing the above inference operation on the Top-K structural templates, the set of virtual attack subchains corresponding to the current node can be obtained. ,Right now .

[0054] Figure 7 A specific embodiment of the virtual attack subchain is shown. The structure shows the virtual target nodes in the virtual attack subchain. It does not point to a real physical identifier ID. Its function is to serve as a matching template in the verification process based on physical dependency matching rules, which will be described later. It filters out attack subchains with sufficiently credible physical causal relationships from various constrained search results of real data, thereby achieving the binding of real nodes. This method of using the semantic information of the inferred virtual target node as a matching template to filter out verifiable target nodes from real data can effectively avoid the problem that the scheme of directly using the inference result as the real node cannot effectively control the illusion of the inference model.

[0055] In some embodiments, such as Figure 4 , Figure 5 or Figure 6 As shown, the steps of "using attack primitives existing in the attack pattern knowledge base and conforming to the environmental characteristics of the current node as inference templates to infer virtual target nodes" and "using only the node semantic attribute information obtained from inference to fill virtual target nodes to generate virtual attack subchains" can be implemented by calling the virtual attack subchain inference module.

[0056] b. Select trusted attack subchains based on physical matching verification. The virtual attack subchain reflects the most likely attack actions and target node attributes that match the current node's scenario. Therefore, it can be used as a "matching template." Based on pre-determined physical dependency matching rules, the candidate attack subchains retrieved from real data are physically matched and verified. If a candidate attack subchain passes the physical matching verification, it means that it has a verifiable causal relationship with the inference result. Furthermore, since it comes from real data and has a real and uniquely bindable physical identifier ID, it can be verified as a trusted attack subchain, thus enabling verifiable link extension starting from the current node.

[0057] In some embodiments, for a current node, the following steps can be used to perform a multi-constraint search on the structured search space and physically match and verify the search results with the virtual subchains to filter out trustworthy attack subchains: The first step is to perform pattern validity verification on the virtual attack subchain. And remove any failed virtual attack subchains.

[0058] The pre-concept pattern validity check is used to verify the virtual attack subchains derived through reasoning. Does it satisfy the physical boundaries of the system itself, including port range, operating system path syntax, etc., if a virtual attack subchain... If the virtual target node is within the physical boundary of the system body, such as its port range being valid and its operating system path being real, it can be used for subsequent physical matching verification. Otherwise, it means that the virtual attack subchain does not have physical feasibility, so it can be directly removed and no longer used as a matching template for physical matching verification.

[0059] The second step involves searching for alternative attack subchains from the structured search space based on multiple constraints.

[0060] Specifically, for virtual attack subchains that have passed pattern validity verification, under multiple constraints, the attack actions of the virtual attack subchain are precisely matched by action type. At the same time, the target node type precisely matches the type of the virtual attack subchain, i.e. As a hit condition, a directed edge and target node satisfying the hit condition are searched in the structured search space to obtain candidate attack subchains. All search hits constitute a set of candidate attack subchains. .

[0061] Optionally, multiple constraints may include one or more of the following: Spatial constraints are used to limit the search scope of the target node to the set of host nodes that the current node can reach in the structured search space.

[0062] The time constraint is used to limit the timestamp of the target node to an adaptive query time window centered on the timestamp of the current node. The length of the time window can be adaptively adjusted based on the statistical results of benign samples. For example, the radius of the time window can be set to the 99th percentile of the maximum normal latency of the system obtained based on the statistics of the most recent 15 million benign I / O samples.

[0063] Data source constraints are used to restrict the source of the target node to multiple independent monitoring data sources that match the virtual target node type and have not been tampered with. middle.

[0064] Figure 8 The diagram illustrates the structure of a candidate attack subchain obtained through a multi-constraint search in a specific embodiment. As can be seen, this subchain originates from a real data source, and therefore its target node's physical identifier ID and semantic attribute information attr both have real values.

[0065] The third step is to physically match and verify the candidate attack subchains obtained from the search with the virtual subchains in order to filter out the trustworthy attack subchains.

[0066] Specifically, using virtual attack subchains To match the template, based on its virtual target node type Select the corresponding physical dependency matching rule For each candidate attack subchain obtained from the search Perform physical dependency checks To determine in the structured search space Does a virtual attack subchain exist within the spatiotemporal window near the current node? Real-world records that match the described attack behavior at the level of physical causal dependencies will identify candidate attack subchains that pass the physical dependency check as trusted attack subchains. This check can be represented as: The dependency rules are as follows for different node types: (1) For network nodes, the following physical dependency rules can be used. Verify network quintuple consistency and time series compliance: , in, It is a network 5-tuple (source IP, source port, destination IP, destination port, protocol). This is the network connection timing tolerance threshold. , These represent the duration of actions for the virtual attack subchain and the alternative attack subchain, respectively. For network nodes, the duration of the action is the same as the duration of the network connection action.

[0067] (2) For file nodes, the following physical dependency rules can be used. Verify the consistency of the mapping from file path to inode number and the consistency of host ownership: , in, The file path of the virtual target node. The inode number of the real file node in the candidate attack subchain. This indicates that the file path of the virtual target node can be mapped to the real file node in the alternative attack subchain. ; This indicates that the virtual target node and the real file node belong to the same host (the same host).

[0068] (3) For process nodes, the following physical dependency rules can be used. Verify the consistency of the image path and the running user, the compliance of the process startup sequence, and the consistency of the parent-child process chain.

[0069] , in, This indicates that the virtual target node is completely identical to the real node in the alternative attack subchain in terms of its mirror path (i.e., the executable file path) and the running user (User). This refers to the process startup time inferred from the virtual attack subchain. The actual startup time of the real processes in the alternative attack subchain. This is the tolerance threshold for process startup timing; The parent process chain (PPID chain) of the real process is consistent with the expectation, that is, the parent-child relationship of the process conforms to the causal logic of the attack chain.

[0070] (4) For registry nodes, the following physical dependency rules can be used. Verify key path consistency and timing compliance: , in, For the registry key path of the virtual attack subchain, The actual registry key path in the alternative attack subchain; This is the tolerance threshold for registry operation timing.

[0071] For a virtual attack subchain, if there is a candidate attack subchain that can be verified by physical matching, it means that there is a record in the real observation and tracing graph that has a causal relationship. Obviously, this candidate attack subchain is a physically verified trusted attack subchain. Since its target node has real physical identification ID information, the link extension of the verification of the current node is realized.

[0072] Figure 9 The figure illustrates the structure of a trusted attack subchain determined by physical matching verification in a specific embodiment. Preferably, a "physical anchor" or other confirmatory identifier can be added to the target node of the trusted attack subchain to indicate that the node has a real physical identifier ID and is verifiable.

[0073] In addition, such as Figure 9 As shown, in addition to carrying a confirmation flag, the selected trusted attack subchains can also carry bias cost and unverified hop count information. Obviously, since the trusted attack subchains have been physically verified, their bias cost and unverified hop count can both be set to 0.

[0074] The selected trusted attack subchains can be directly connected to the current node, thereby enabling link extension operations from the current node. Figure 10 The diagram illustrates a specific embodiment of using a selected trusted attack subchain to extend the link from the current node. It can be seen that since the bias cost and unverified hop count of the trusted attack subchain are both 0, they will not affect the accumulated bias and accumulated unverified hop count of the entire link.

[0075] For a virtual attack subchain, if two or more candidate attack subchains that can pass the physical dependency matching check can be selected, they can be used as trusted attack subchains respectively. In this way, a fork can be formed from the current node, generating two or more candidate attack links for the next link extension. Alternatively, the semantic distance or semantic similarity between these candidate attack subchains and the virtual attack subchain can be calculated separately, and only the candidate attack subchain with the closest semantic distance (highest semantic similarity) can be used as a trusted attack subchain.

[0076] In some alternative embodiments, such as Figure 4 , Figure 5 or Figure 6 As shown, the steps described above, "performing a multi-constraint search on the structured search space and physically matching and verifying the search results with the virtual subchains to filter out trusted attack subchains," can be implemented by calling the trusted attack subchain filtering module.

[0077] c. Determine whether to continue link extension based on inference results based on the deviation cost assessment. If a virtual attack subchain cannot be selected, i.e. no candidate attack subchain that can pass the physical dependency matching check is selected, it indicates that the candidate attack subchain retrieved within the spatiotemporal constraint window of the current node in the untampered data has a weak causal relationship with the virtual attack subchain. In this case, the virtual attack subchain can be used as an inference attack subchain and its bias cost can be evaluated to determine whether to continue the link extension starting from the current node.

[0078] In some embodiments, bias cost evaluation for any virtual attack subchain includes the following steps: The first step is to determine the comprehensive deviation cost based on the contextual semantic deviation and time delay penalty value of the virtual target node.

[0079] Specifically, the semantic attributes and associated action edges of the virtual target node can be converted into normalized text, and the first feature vector can be extracted. The high-dimensional feature vector of the most relevant attack primitive retrieved is used as the second feature vector. Calculate the cosine distance between the two as the semantic deviation from the context: .

[0080] Specifically, the time difference between the virtual target node and its parent node can be calculated. And based on the log-normal cumulative distribution function fitted by historical benign I / O samples. Calculate the time delay penalty value: , in, This indicates that in the case of violating causality, the cost deviation of the virtual attack subchain will become infinite. In actual calculation, it can be set to a large number that is far greater than the deviation cost threshold that can trigger the truncation of the chain.

[0081] Calculated and Then, the baseline distribution data of the task can be reconstructed based on historical normal logs. and Perform Z-score normalization to obtain the normalized result. and Then, the comprehensive deviation cost is obtained by weighted summation: , in, , These are the weighting coefficients optimized based on a benign baseline grid search.

[0082] The second step is to use the virtual attack subchain as the inference attack subchain to extend the link of the current node, and determine whether the extended current candidate attack link meets the dynamic budget admission conditions. If it does, the extended current candidate attack link is retained as the candidate attack link for the next round of link extension; otherwise, the extended current candidate attack link is directly deleted.

[0083] Figure 11 The diagram illustrates a structure using a virtual attack subchain as the inference attack subchain. It shows that the target node in this inference attack subchain lacks a definite real physical identifier (ID), possessing only semantic information (attr) generated through inference. Furthermore, the target node carries an "unverified" or similar confirmatory identifier. Its bias cost is the comprehensive bias cost calculated earlier. If the number of jumps is not verified, it will be set to 1.

[0084] Figure 12 A schematic diagram illustrating link extension using this inference attack subchain is shown. It can be seen that after extension, The accumulated bias cost will be added to the expanded current candidate attack chain. In addition, the cumulative number of unverified hops of the current candidate attack link after expansion. Also add 1. At this point, the following formula can be used to determine whether the dynamic budget admission condition is met after using the inference attack subchain for link expansion: , in, The preset length penalty factor, for example, can be taken as... , The maximum search cost budget can be set, for example, to the 99th percentile of the cost distribution obtained from statistics of benign host single-hop log completion tasks.

[0085] If the judgment result is yes, it means that after adding the attack sub-chain obtained by inference, the accumulated cost of the entire chain has not exceeded the acceptable range. At this time, it can still be used as a candidate attack chain for the next round of chain expansion, and there is still a chance to carry out the next round of expansion (if it is not pruned by the bundle width constraint). Otherwise, it means that as the chain is expanded using the inference result, the deviation has accumulated to an unacceptable level, so the continued expansion of the candidate attack chain is stopped directly.

[0086] In some embodiments, such as Figure 4 , Figure 5 or Figure 6 As shown, the above-mentioned "when a trustworthy attack subchain cannot be selected, the virtual attack subchain is used as the inference attack subchain and its bias cost is evaluated" can be implemented by calling the inference attack subchain evaluation module.

[0087] <Output of Threat Hunting and Source Tracing Results> In some optional embodiments, if the attack chain search module can find at least one complete attack chain from the starting anchor point to the ending anchor point for a threat hunting and attribution request, it can directly output the complete attack chain found.

[0088] In some optional embodiments, if the attack chain search module fails to find any complete attack chain in response to a threat hunting and tracing request, a security termination mechanism can be triggered and a search failure message, such as the INSUFFICIENT_EVIDENCE status code, can be output. Alternatively, the incomplete attack chain that has been searched can also be output.

[0089] In some alternative embodiments, such as Figure 13 As shown, the system also includes a layered encapsulation module, which is used to encapsulate the search results in layers based on the confidence level of each attack subchain of the attack chain obtained by the search.

[0090] For example, all attack subchains containing nodes marked with "physical anchoring" can be extracted as physically trusted links to construct a physical verification subgraph. This results in the encapsulation and output of highly credible objective physical evidence data. For example, it is possible to extract the attack subchains containing nodes marked "unverified" and filter them out. A heuristic inference subgraph is constructed based on attack subchains that do not exceed a preset deviation threshold. The clues are marked as auxiliary investigation clues, and then encapsulated and output after adding the comprehensive deviation value and the matching attack primitives.

[0091] Figure 14 The figure illustrates the specific flow of the attack link search module performing step 200, i.e., the heuristic path search algorithm based on neural symbolic architecture, in a specific embodiment. As shown in the figure, the algorithm execution process consists of initialization operations and cyclic link expansion operations.

[0092] Initialization operation: The initialization process includes generating a set of candidate attack links. And initialize the candidate attack links, where each candidate attack link has the following format: That is, the current node The generated attack chain Cumulative deviation cost and cumulative unverified jump count Clearly, during the initialization phase, There is only one unexpanded path consisting solely of the initial anchor point. The constructed candidate attack chain, namely .

[0093] Initialization also includes initializing algorithm hyperparameters, such as setting the beam search width. (Preferred, This means that in each round of link expansion operations, there are no more than 6 candidate attack links to avoid an explosive growth of links in the later stages of the search. (This refers to the number of attack patterns retrieved.) (like ), retrieval neighborhood hop count (like ), length penalty factor (like Maximum search cost budget (such as the 99th percentile value calibrated at a benign baseline).

[0094] Link extension operation: After the initialization operation is completed, the link expansion operation can be performed cyclically. In the embodiments of this application, the search framework for the link expansion operation is beam search, that is, after each round of link expansion operation, only w links are retained to enter the next round of expansion (w is the preset beam width).

[0095] refer to Figure 14 In each round of link expansion operations, candidate attack links are extracted sequentially from the candidate attack link set as the current candidate attack link, and then it is determined whether the current node of the current candidate attack link is the termination anchor point. If the judgment result is yes, it means that the candidate attack link has been reconstructed into a complete attack link. Therefore, it is directly added to the search results set, and it is judged whether all candidate attack links in this round have been traversed. If the judgment result is yes, the link expansion operation in this round ends.

[0096] If the current node is not the terminating anchor point Then, the attack chain search module calls the virtual attack subchain inference module, and by performing the aforementioned attack pattern search and constrained inference, obtains the set of virtual attack subchains for the current node. Then, each virtual attack subchain is extracted sequentially, and the trusted attack subchain matching module is called to perform the aforementioned physical matching verification operation. If a trusted attack subchain can be selected through the physical matching verification operation, the trusted attack subchain is directly used to extend the current attack link, and the extended attack link is placed into the candidate attack link set for the next round, and then the processing of the next virtual attack subchain begins. If a trusted attack subchain cannot be selected, the inference attack subchain evaluation module is called to evaluate whether the cumulative cost of extending the link to the current node using the virtual attack subchain as the inference attack subchain meets the dynamic admission conditions. If the evaluation passes, the extended attack link is added to the candidate attack link set for the next round, and then the processing of the next virtual attack subchain begins. Otherwise, the new attack link extended using the inference attack subchain is directly removed.

[0097] Once all virtual attack subchains have been processed, the link expansion of a candidate attack chain is complete. Then, other candidate attack chains are selected from the candidate attack chains, and the above inference-matching-bias evaluation operation based on neural symbolic architecture is repeated until all candidate attack chains in this round of link expansion operation have been traversed.

[0098] At this point, the current round of operations ends, and it is determined whether any attack links have been added to the candidate attack link set for the next round. If the candidate attack link set for the next round is empty, it means that during the current round of expansion operations, either a complete attack link has already been added to the search results set, or the accumulated deviation is too large and exceeds the acceptable range, making it unnecessary to continue link expansion. At this point, the search algorithm can end. If a complete attack link exists in the search results set, the complete attack link is output as the threat hunting and tracing result, or the layered encapsulation method described above is used to output highly credible objective physical evidence and auxiliary investigation results respectively. Otherwise, information that no complete attack link was found can be output, and the attack link fragments that have been searched can be output as needed.

[0099] If the next round of candidate attack links is not empty, then the number of attack links within it is further determined; if the number exceeds a preset threshold... When searching for the beam width, the results can be filtered based on the ranking of the accumulated bias costs of each attack link. If the number of attack links does not exceed a certain limit, then all of them will be used in the next round of link expansion. It can be understood that before the next round of operation, the results of this round of expansion will be used to update the candidate attack link set, including updating the current node to the target node of the expanded attack sub-chain, and updating the accumulated deviation cost and accumulated unverified hop count.

[0100] <Specific Implementation Example 1> This embodiment builds a threat hunting and tracing system based on neural symbolic architecture in a real experimental environment and verifies its threat hunting and tracing effect in anti-forensic scenarios.

[0101] A. Experimental Environment Experimental hardware environment: 3-node KVM cluster, each node is configured with 2×Intel Xeon Gold 6248R CPUs, 512GB memory, and 2×NVIDIA A100 80GB GPUs; Experimental software environment: Python 3.10, PyTorch 2.1, FAISS 1.7.4, OpenAI GPT-4o API, Anthropic Claude 3.5 Sonnet API, Meta Llama-3-70B-Instruct, Alibaba Qwen2.5-72B-Instruct; General parameters for the large model used for inference: temperature coefficient τ=0.2, top_p=0.9, maximum generation length 2048 tokens.

[0102] B. Experimental Dataset The experiments used one publicly available standard dataset (DARPA OpTC, D1) and one self-built dataset based on real APT attacks (APT-Eval-Trace, D2), covering multiple operating systems including Windows, Linux, and FreeBSD, and containing more than 40 real APT attack scenarios. All publicly available datasets were injected with anti-forensic perturbations consistent with real APTs using the EVADEKIT tool, while the self-built dataset contained natively executed anti-forensic operations.

[0103] Table 2 summarizes the dataset information: Table 2 C. Implementation Status Table 3 shows the search algorithm parameters: Table 3 Threat model assumptions: During the experiment, it is assumed that at least one independent monitoring source survives at the relevant operating system layer; attackers can perform user-space anti-forensic operations such as log erasure, PID spoofing, and fileless execution, but cannot simultaneously tamper with all independent monitoring tools.

[0104] Verification method: Attack family isolated cross-validation (LOFO) is adopted. The attack families in the test set do not appear in the training set and attack pattern knowledge base to avoid inflated results due to data leakage.

[0105] E. Selection of baseline for comparison Baseline Comparison: Four of the most representative APT attack attribution techniques in the prior art were selected as baseline comparisons and compared with the results of the threat hunting and attribution system provided in this application. Table 4 summarizes the baseline comparisons. Table 4 Baseline Adaptation Notes: KAIROS and NODLINK were originally designed for node-level anomaly detection. This experiment adapts them for attack link reconstruction, extracting the highest-confidence connected subgraph from the alarm anchor point to the target node as the prediction result. All baselines use the same input data, anti-forensic perturbations, and evaluation criteria to ensure fairness in comparison.

[0106] F. Evaluation Indicators Five evaluation metrics were used to comprehensively assess the source tracing performance. The evaluation basis was determined as follows: First, a causal equivalence relation was defined, and for any prediction edge... If a real attack edge exists If the following causal equivalence conditions are met, then... The result is a true positive (TP): , The causal equivalence conditions specifically include: (1) the same underlying state transition type (process creation, file writing, network connection, registry modification); (2) the causal equivalence execution subject (confirmed through PPID chain or PID alias resolution); and (3) the timestamp difference does not exceed 1 second.

[0107] 1) Precision (P): The proportion of edges predicted as attack edges that are actually attack edges. It measures the reliability of the result. The mathematical formula is: , FP (False Positive) represents the number of false positive edges, i.e., the number of benign edges that are incorrectly predicted as attack edges.

[0108] 2) Recall (R): The proportion of real attack edges that are correctly predicted, measuring the completeness of the results. The mathematical formula is: , FN (False Negative) represents the number of false negative edges, i.e., the number of real attack edges that were not predicted.

[0109] 3) F1 score: The harmonic mean of precision and recall, which comprehensively measures the source tracing performance. The mathematical formula is: , 4) Path Hallucination Rate (PHR): The proportion of generated attack paths containing edges that violate the physical laws of the operating system. It measures the severity of false inference in large models. The mathematical formula is: , in This is the set of all attack paths generated by the system. This is a physical rule validation function for the operating system, returning True if and only if the edge... Violation of any of the following rules: (1) Reversed causal order (child node time is earlier than parent node); (2) Violation of system ontology (e.g., non-privileged process injects protected process); (3) Incorrect resource access permissions; (4) Physically unreachable network connection.

[0110] 5) Budget exhaustion rate: The proportion of investigations that trigger the safety termination mechanism due to insufficient evidence out of the total number of investigations, which measures the conservatism of the system.

[0111] G Results Analysis G-1 overall traceability performance comparison: In a standard anti-forensic scenario with 30% target log erasure, the attack family isolated cross-validation (LOFO) method is used to compare the overall performance of this invention with four mainstream source tracing techniques on four datasets. Table 5 shows the overall source tracing performance comparison results: Table 5 As can be seen, the system and method provided in this application achieve an F1 score of 85.7% on the public dataset D1 and 86.8% on the self-built dataset D2, both of which are 17.1 percentage points higher than the average level of learning-based tracing techniques and 26.7 percentage points higher than the average level of tracing techniques based on large language model agents. The average path illusion rate of this invention on the two datasets is 6.7%, which is 54.6 percentage points lower than the average level of tracing techniques based on large language model agents. The accuracy of rule-based deterministic tracing techniques is higher than 98%, but the recall rate is lower than 36% in the 30% log erasure scenario.

[0112] G-2's ability to resist counter-forensic interference verification: On the publicly available dataset D1 (DARPA OpTC), four different target log erasure scenarios with 10%, 30%, 50%, and 70% were simulated to test the performance variation of this application. The results are listed in Table 6. Table 6 As can be seen, when the log erasure rate increases from 10% to 70%, the accuracy of this invention decreases from 93.1% to 84.0%, a decrease of 9.1 percentage points; in the extreme log erasure scenario of 70%, 95.7% of the investigations trigger the security termination mechanism and return the INSUFFICIENT_EVIDENCE status code.

[0113] G-3 Key Module Performance Verification On the D2 (APT-Eval-Trace) dataset, core modules involving neural symbolic architectures were removed sequentially, and performance changes were tested. The results are listed in Table 7. Table 7 As can be seen, after removing the physical verification-related modules, the path illusion rate increased from 6.4% to 28.6%, proving that physical hard constraints are the core to preventing false inference; after removing the bias cost assessment and dynamic budget admission-related modules, the accuracy decreased to 45.8%, proving that dynamic budget is the key to controlling inference drift. This shows that all core components contribute significantly to the performance of this invention.

[0114] G-4 Inference Model Irrelevantness Verification The performance of this application was tested on the D1 (DARPA OpTC) dataset using large models from different vendors as generators. The results are listed in Table 8. Table 8 As can be seen, when using different large models as generators, the accuracy rate of this application remains above 85%, and the path illusion rate is below 12%. The performance difference mainly comes from the semantic understanding ability of the large models themselves, which proves that the neural symbolic architecture of this application has broad versatility and does not depend on the large model services of specific vendors.

[0115] G-5 Operational Efficiency and Cost Assessment On the D2 (APT-Eval-Trace) dataset, the average runtime, token consumption, and currency cost of a single survey are statistically analyzed, and the results are listed in Table 9: Table 9 As can be seen, the cost per survey in this application is only $0.19, which is 69.3% lower than ReAct and comparable to a single-round GraphRAG; the average running time is 42.8s, which is 76.9% faster than ReAct, meeting the real-time survey needs of enterprise SOCs.

[0116] G-6 Adaptive Countermeasures Verification On the D1 (DARPA OpTC) dataset, the ability of the proposed scheme to resist targeted attacks, including NetFlow injection, ETW provider patching, and calibration drift, was tested. The results are listed in Table 10. Table 10 As can be seen, even under the most severe combined attacks, the path illusion rate of the proposed solution is only 18.4%, far lower than the 61.5% of the ReAct method based on large model agents; all attacks mainly affect the recall rate rather than the precision rate, indicating that it always maintains a high reliability of results.

[0117] G-7 Hyperparameter Sensitivity Analysis On the D1 (DARPA OpTC) dataset, the sensitivity of this invention to core hyperparameters was tested, and the results showed that the weighting coefficients ( , When the value varies within the range of [0.3, 0.9]², the F1 value fluctuates by no more than 2 percentage points; length penalty factor The optimal F1 score of 85.7% was achieved when the F1 value was 0.25. When the value is 0, the F1 value is 80.3%; The F1 score is 81.6% at →∞; maximum search cost budget The optimal balance is achieved when the value is set at the 99th percentile. Setting it too tight will lead to a decrease in recall, while setting it too loose will lead to an increase in hallucination rate.

[0118] The specific embodiments of this application have been described in detail above. For those skilled in the art, several improvements and modifications can be made to this application without departing from the principle of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A threat hunting and tracing system based on a neural symbolic architecture, characterized in that, include: A structured search space is used to store observation source maps that conform to time-series causal constraints, generated based on raw monitoring data; An attack pattern knowledge base is used to store attack pattern graphs. The attack pattern graphs are consistent with the observation and tracing graphs in terms of node types and edge semantics, including attack primitives that represent various standardized attack patterns. The attack chain search module, in response to threat hunting and attribution requests, searches for complete attack chains from the starting anchor point to the ending anchor point by executing a heuristic search algorithm based on neural symbolic architecture. In each chain expansion operation of the algorithm, a new trusted attack sub-chain or inferred attack sub-chain is generated starting from the current node of the current candidate attack chain based on the following steps: Using attack primitives that exist in the attack pattern knowledge base and conform to the environmental characteristics of the current node as inference templates, virtual target nodes are inferred. Virtual target nodes are populated using only the semantic attribute information of nodes obtained through inference to generate virtual attack subchains; A multi-constraint search is performed on the structured search space, and the search results are physically matched and verified with the virtual subchains to filter out the trustworthy attack subchains. When a trustworthy attack subchain cannot be identified, the virtual attack subchain is used as the inference attack subchain, and its bias cost is evaluated.

2. The threat hunting and tracing system based on neural symbolic architecture according to claim 1, characterized in that, The observation source map This refers to a banded causal graph conforming to the system ontology specification, obtained by performing loop removal processing on multi-source heterogeneous monitoring data in the enterprise environment according to time-series causal constraints. for The set of nodes, including process nodes, file nodes, network nodes and registry nodes, where each type of node includes a physical identifier ID field and a semantic attribute field; for A set of directed edges that follow temporal causality constraints. The directed edges in the algorithm are defined as causal transformation structures, which include the source node that initiates the action, the standardized action, the target node toward which the action is directed, and the timestamp of the action.

3. The threat hunting and tracing system based on neural symbolic architecture according to claim 1, characterized in that, The attack pattern diagram This refers to a typed causal graph extracted from multi-source threat intelligence data, conforming to the system ontology specification. This refers to the set of nodes involved in various extracted attack actions; For each type of extracted attack action, there are causal edges. Each causal edge represents an attack action that points from the source node to the target node.

4. The threat hunting and tracing system based on neural symbolic architecture according to claim 3, characterized in that, The process of inferring virtual target nodes using attack primitives existing in the attack pattern knowledge base and conforming to the environmental characteristics of the current node as inference templates includes: Using the current node as the source node, extract its The neighboring domain is skipped and serialized into a standard JSON format, mapped to a query vector through the same semantic encoding model, and the top-K primitives with the closest cosine similarity are retrieved from the attack pattern knowledge base based on the query vector, which serve as the attack primitive set adapted to the current scenario. The retrieved attack primitives are used as structural templates, combined with the context, and then input into the inference model. The inference model generates virtual target nodes under the constraints of prior attack patterns that conform to the current scenario.

5. The threat hunting and tracing system based on neural symbolic architecture according to claim 3, characterized in that, The process of performing a multi-constraint search on the structured search space and physically matching the search results with the virtual subchains to filter out trustworthy attack subchains includes: Perform pattern validity verification on the virtual attack subchains and remove those that fail. Based on multiple constraints, alternative attack subchains are searched from the structured search space; The search-obtained candidate attack subchains are physically matched and verified with the virtual subchains to filter out the trustworthy attack subchains.

6. The threat hunting and tracing system based on neural symbolic architecture according to claim 5, characterized in that, The multiple constraints include at least one of the following: Spatial constraints are used to limit the search scope of the target node to the set of host nodes that the current node can reach in the structured search space; The time constraint is used to restrict the timestamp of the target node to an adaptive query time window centered on the timestamp of the current node, and the length of the time window can be adaptively adjusted according to the statistical results of benign samples. Data source constraints are used to restrict the source of the target node to multiple independent monitoring data sources that match the virtual target node type and have not been tampered with.

7. The threat hunting and tracing system based on neural symbolic architecture according to claim 5, characterized in that, The step of physically matching and verifying the candidate attack subchains obtained from the search with the virtual subchain to filter out trustworthy attack subchains includes: Using the virtual attack subchain as a matching template, the corresponding physical dependency matching rule is selected according to the type of its virtual target node. Physical dependency checks are performed on each candidate attack subchain obtained from the search, and the candidate attack subchains that pass the physical dependency check are determined as trusted attack subchains.

8. The threat hunting and tracing system based on neural symbolic architecture according to claim 3, characterized in that, For any virtual attack subchain, when a trustworthy attack subchain cannot be identified, the virtual attack subchain is treated as an inference attack subchain and its bias cost is evaluated, including: The comprehensive deviation cost is determined based on the contextual semantic deviation and time delay penalty value of the virtual target node; The virtual attack subchain is used as the inference attack subchain to extend the link of the current node. It is determined whether the extended current candidate attack link meets the dynamic budget admission conditions. If it does, the extended current candidate attack link is retained as the candidate attack link for the next round of link extension. Otherwise, the extended current candidate attack link is directly deleted.

9. The threat hunting and tracing system based on neural symbolic architecture according to claim 8, characterized in that, The dynamic budget admission criteria are as follows: , in, For the expanded current candidate attack chain The cost of accumulated bias This represents the cumulative number of unverified hops in the expanded current candidate attack path. As a length penalty factor, Budget for maximum search cost.

10. The threat hunting and tracing system based on neural symbolic architecture according to claim 1, characterized in that, Also includes: The search space management module is used to establish and manage the structured search space. The knowledge base management module is used to establish and manage the attack pattern knowledge base. The virtual attack subchain reasoning module is called by the attack chain search module to implement the steps of reasoning virtual target nodes using attack primitives that exist in the attack pattern knowledge base and conform to the environmental characteristics of the current node as reasoning templates, and filling virtual target nodes with only the node semantic attribute information obtained by reasoning to generate virtual attack subchains. The trusted attack subchain matching module is called by the attack chain search module to implement the steps of performing multi-constraint search on the structured search space and physically matching and verifying the search results with the virtual subchains in order to filter out trusted attack subchains. The inference attack subchain evaluation module is called by the attack chain search module to perform the step of using the virtual attack subchain as the inference attack subchain and evaluating its bias cost when a trusted attack subchain cannot be found.

11. The threat hunting and tracing system based on neural symbolic architecture according to claim 1, characterized in that, Also includes: The layered encapsulation module is used to encapsulate the search results in layers based on the confidence level of each attack subchain of the attack chain obtained from the search.

12. A threat hunting and attribution method based on a neural symbolic architecture, using the system described in claim 1 for threat hunting and attribution, characterized in that, include: In response to threat hunting and attribution requests, attribution boundaries are determined, including attribution start and end anchor points; A heuristic attack chain search algorithm based on neural symbolic architecture is executed to search for each complete attack chain from the starting anchor point to the ending anchor point. The complete attack chain includes each attack sub-chain connected sequentially between the starting anchor point and the ending anchor point.