A Security Incident Assessment Method Based on Minimum Sufficient Evidence Subgraph
By constructing a minimum sufficient evidence subgraph and a constraint judgment method, the fragmentation and miscorrelation problems of multi-source alarms in network attacks are solved, achieving efficient and accurate security event judgment and simplifying the analysis process for operators.
Patent Information
- Application Number
- CN202611577397.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-13
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies, when dealing with cyberattacks, suffer from fragmented multi-source alerts, excessively large candidate evidence graphs that overwhelm critical behaviors, weak contextual misassociations, and a disconnect between model analysis conclusions and original records, making it difficult to meet the needs of efficient and secure operations.
By constructing a minimum sufficient evidence subgraph, using explicit connection keys and hierarchical relationship generation rules, a minimum sufficient evidence subgraph covering key behaviors is extracted. A general large model is then used for constraint assessment and independent evidence verification to generate intuitive and verifiable assessment results.
It effectively removes duplicate transactions and irrelevant context, reduces false associations, improves the accuracy of judgment, shortens the analysis cycle, and ensures the reliability and verifiability of conclusions.
Smart Images

Figure CN123074844A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of network security detection and analysis technology, and in particular relates to a security event assessment method based on the minimum sufficient evidence subgraph. Background Technology
[0002] The complexity and stealth of cyberattacks continue to rise. Advanced Persistent Threats (APTs) and ransomware attacks typically involve multiple stages, including reconnaissance, initial access, privilege escalation, lateral movement, and data infiltration. Attacks leave scattered log traces at multiple levels, including endpoints, networks, identities, and applications. Security Operations Centers (SOCs) receive massive amounts of alerts and data daily from multiple systems, including intrusion detection systems, endpoint detection and response (EDR), host auditing, identity authentication, web services, and network traffic. Analysts need to reconstruct the complete attack chain from fragmented, multi-source data to determine whether alerts constitute a real threat and to define the scope of action. Alert fatigue and insufficient analysis efficiency have become core issues restricting the effectiveness of security operations.
[0003] Current alarm aggregation methods group alarms and logs together based on simple criteria such as time windows, network addresses, and host identifiers. However, proximity in time or shared addresses does not necessarily equate to causal relationships, easily leading to the incorrect merging of unrelated activities and generating numerous false associations. Furthermore, these methods fail to clarify the chronological order of actions between records, thus not substantially reducing the processing burden. Another approach involves constructing a large-scale event graph from the logs and using graph search or graph neural networks to identify anomalies. However, candidate relationships grow rapidly with data volume, resulting in a graph containing numerous duplicate transactions and irrelevant relationships, easily obscuring critical actions. Outputting a large and comprehensive graph makes it difficult for operations personnel to directly identify the attack origin, key operations, and affected objects, and also hinders the retrieval of relationships from the graph back to the original logs.
[0004] Another approach involves flattening the retrieved logs and then automatically analyzing them using rule engines, statistical models, or language models. However, flattening the input lacks structured information such as entity switching, transaction continuation, and cross-data source relationships. Key evidence in long contexts can easily be buried under ordinary records. Furthermore, the model output lacks a verifiable correspondence with the original logs, making the conclusions difficult to audit and verify. Without automation, operations personnel must manually search and analyze logs across different sources. Reconstructing an attack often requires reviewing a large number of original records, resulting in long processing times, high reliance on experience, inconsistent and unintuitive outputs, and difficulty meeting the timeliness requirements of scenarios with large alert volumes and fragmented contexts. Summary of the Invention
[0005] The purpose of this invention is to provide a security event analysis method based on the minimum sufficient evidence subgraph, which solves the problems of fragmented multi-source alarms, excessively large candidate evidence graphs leading to the submergence of key behaviors, weak contextual misassociations, and the disconnect between model analysis conclusions and original records, making them unauditable.
[0006] To achieve the above objectives, the present invention provides a security event assessment method based on a minimum sufficient evidence subgraph, comprising the following steps: S1. Obtain multi-source logs and alarms involved in the security event to be analyzed, and represent them as a unified set of event nodes with additional original record references. Use explicit join keys to expand the context and recall candidate events to obtain an expanded event set. S2. Execute the hierarchical relationship generation rule on the extended event set to generate a hierarchical relationship set containing the direct operation layer, the complete transaction layer, the multi-event behavior layer and the weak context layer. Construct a candidate evidence graph with event nodes as vertices and hierarchical relationships as edges. S3. Extract the minimum sufficient evidence subgraph from the candidate evidence graph that covers key behaviors, has traceable relationships, and has closed paths. Assign unique numbers to the nodes and relationships of the minimum sufficient evidence subgraph, generate evidence credentials, and form a set of evidence numbers that can be cited. S4. Input the minimum sufficient evidence subgraph, the original records it references, and the set of evidence numbers that can be referenced into the constraint judgment module. Use the general large model to perform constraint judgment and generate judgment results. Perform independent evidence verification on the judgment results. If the verification passes, output the judgment results directly. If the verification fails, return to the constraint judgment module for secondary reasoning. If it still fails, output insufficient evidence.
[0007] Preferably, in step S1, alarm event nodes are extracted from alarms to be analyzed, context logs are parsed from multi-source logs and normalized into a unified event structure, context log event nodes are constructed, and original record references including log source, record location, occurrence time, and original content summary are attached to each event node to obtain the unified event node set; the explicit connection key includes session number, request number, network flow number, process number, parent and child process identifier, account switching record, network 5-tuple, and process-socket mapping.
[0008] Preferably, in S2, the hierarchical relationship generation rule is as follows: for a single event record, event record pair, and event record sequence in the extended event set, first determine whether the value requirements of the matching relationship rule are met in the time, subject, object, action, and data source fields; then determine whether the association requirements of the matching relationship rule are met in the connection key, session, transaction, and behavior sequence; finally determine whether the event pair has a supporting event number and original record reference. When all three judgments pass, the rule number of the matching relationship rule and the fields involved in the judgment are written into the relationship record and the relationship is generated according to the corresponding semantic level. When any judgment fails, only a weak context layer relationship is generated.
[0009] Preferably, in S2, the direct operation layer relationship represents an operation that can be directly verified for a single record; the complete transaction layer relationship represents a closed transaction consisting of a request and response, a connection start point and a connection end point, and an authentication operation and authentication result; the multi-event behavior layer relationship represents a sequence of behaviors formed by connecting the direct operation layer relationship and the complete transaction layer relationship; and the weak context layer relationship represents a candidate relationship formed by shared address, shared host, or time proximity.
[0010] Preferably, the construction of the candidate evidence graph in S2 is specifically as follows: the event nodes in the extended event set are used as vertices, the relationships in the hierarchical relationship set are used as edges, and the candidate evidence graph is organized according to the data source view; the relationships within the same data source view form a subgraph within the view, and the relationships across data source views are used to connect the subgraphs within the view.
[0011] Preferably, the data source views include four types: host and authentication views generate relationships based on process ID, parent-child process ID, session ID, account switching records, and command execution records; web views close the same request transaction using request ID, connection ID, client address, service ID, and resource path; network views close connections using network flow ID or network 5-tuple, and close domain name resolution requests and responses using transaction ID; cross-view association uses request ID, trace ID, network flow ID, session ID, or process-socket mapping that can appear in two views simultaneously.
[0012] Preferably, the minimum sufficient evidence subgraph extraction process in S3 is as follows: First, determine the ordered critical path corresponding to the event to be assessed. The ordered critical path starts from the event seed, passes through key behaviors, and reaches the affected object. The key relationships on the path belong to one or more of the following categories: direct operation layer relationships, complete transaction layer relationships, and multi-event behavior layer relationships. Each key relationship records a supporting event number and an original record reference. Paths composed entirely of weak context layer relationships do not satisfy path closure. While maintaining key behavior coverage, relationship retrospectiveability, and evidence closure, unnecessary relationships are deleted from the candidate evidence graph to obtain the minimum sufficient evidence subgraph. A unique number is assigned to each node and each relationship in the minimum sufficient evidence subgraph, generating evidence credentials including subgraph number, node number, relationship number, ordered critical path, supporting event number, and original record reference. The node number and relationship number are extracted from the evidence credentials to form the set of allowed citation evidence numbers.
[0013] Preferably, the minimum sufficient evidence subgraph is selected from the candidate evidence graphs according to the following objective function: ; ; ; ; Where G is the candidate evidence graph, H is the extracted critical path subgraph, and G* is the minimum sufficient evidence subgraph; Coverage(H) is the critical behavior coverage rate, representing the proportion of the number of critical behavior types covered by subgraph H to the number of critical behavior types specified in the rule template; Trace(H) is the relationship traceability rate, representing the proportion of nodes and relationships in subgraph H that have original record references or supporting event references to the total number of nodes and relationships in subgraph H; Redundancy(H) is the redundant relationship ratio, representing the proportion of duplicate transaction relationships and relationships in subgraph H that, after deletion, do not affect the closure of evidence to the total number of relationships in subgraph H; n cover (H) represents the number of key behavior types covered by subgraph H, n required n represents the number of key behavior types specified in the rule template corresponding to the event to be analyzed. trace (H) represents the number of nodes and relationships in subgraph H that have original record references or supporting event references, n. redundant (H) represents the number of duplicate transaction relations in subgraph H and the number of relations that do not affect the closure of evidence after deletion. |V(H)| and |E(H)| represent the number of nodes and the number of relations in subgraph H, respectively. λ is the penalty coefficient for the number of relations. For candidate subgraphs with the same target value, the subgraph with the relation number sorted first is selected.
[0014] Preferably, the data processing flow of the constraint assessment module is as follows: The minimum sufficient evidence subgraph, the original record fragments it references, and the set of allowed evidence numbers are input into the constraint judgment module. The general large language model is used to generate the judgment result. By using one or more of the following methods, such as the evidence number whitelist in the prompt words, structured output mode, JSON Schema constraints, syntax-guided decoding, output tag mask, and probability bias, the range of evidence numbers referenced in the judgment process is restricted, and a structured judgment result including the judgment conclusion, confidence level, reasoning, referenced evidence numbers, and missing evidence is generated.
[0015] Preferably, the specific process of performing independent evidence verification on the assessment results is as follows: Each cited evidence number is checked to ensure it belongs to the set of allowed evidence numbers. The semantic level, rule number, and supporting event number of each number's corresponding relationship are read to determine whether the cited relationship can support the reasoning behavior. It is also checked whether the ordered critical path starts from the event seed, reaches the affected object through the critical behavior, and whether the critical relationships on the path all have supporting events and original record references. If the verification passes, the structured judgment result is output as the final judgment conclusion. If the verification fails, the invalid references and missing evidence are returned to the constraint judgment module for secondary reasoning. The evidence numbers cited in the secondary reasoning are still limited to the set of allowed evidence numbers. After the secondary reasoning, independent evidence verification is performed again. If it still fails, the conclusion is corrected to insufficient evidence and output, along with the critical path, valid evidence, and original record references.
[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The minimum evidence subgraph is directly used as the input for the large language model, replacing the full log flat input, removing duplicate transactions and irrelevant context, avoiding irrelevant records from interfering with the model's judgment, reducing the illusion that the model references content that is unrelated to or does not exist, and significantly reducing the input length and token consumption. 2. By organizing multi-source records through explicit join keys and hierarchical relationships, weak contexts are restricted to the candidate recall stage, avoiding the formation of closed attack conclusions based solely on the same address, host, or time window, thus reducing false associations; 3. Retain the ordered critical path and its direct support to enable the model to focus on key behaviors and improve the accuracy of judgment; 4. The minimum evidence subgraph presents the attack origin, key behaviors, and affected objects in a graphical form, which is more intuitive than flat logs. Operations personnel can understand the whole picture of the event without having to go through each original record, which helps to make quick judgments and shortens the judgment cycle. 5. Each basis for the analysis conclusion is checked back to the original log through the evidence number, and missing evidence prompts are output. The conclusions are verifiable, and the operators can check them one by one, which alleviates the black box effect of the conclusions of the large language model. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.
[0018] Figure 2 This is a schematic diagram of the layered evidence diagram construction process of the present invention.
[0019] Figure 3 This is a schematic diagram of the hierarchical evidence relationship of the present invention.
[0020] Figure 4 This is a schematic diagram of the minimum sufficient evidence subgraph extraction process of the present invention.
[0021] Figure 5 This is a schematic diagram of the evidence constraint assessment and independent evidence verification process of this invention.
[0022] Figure 6 This is a schematic diagram comparing the experimental results of the present invention; where (a) is a comparison diagram of the number of relationships between the candidate evidence graph and the minimum sufficient evidence subgraph of the present method; and (b) is a comparison diagram of the judgment effect between the shared entity relationship graph and the minimum sufficient evidence subgraph. Detailed Implementation
[0023] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0024] This invention proposes a security event assessment method based on the minimum sufficient evidence subgraph. This method follows... Figure 1 The process shown sequentially completes the following steps: unified event representation, candidate event expansion, hierarchical relationship generation, candidate evidence graph construction, minimum sufficient evidence subgraph extraction, evidence credential generation, constraint assessment, and independent evidence verification. The following example uses a host privilege escalation security event to illustrate the execution of each step.
[0025] This embodiment receives alarms pending analysis, authentication logs, host audit logs, web access logs, network flow records, asset records, and detector outputs. Alarms pending analysis indicate that a web service host has engaged in privilege escalation behavior; the event metadata provides the alarm number, occurrence time, host identifier, and web service account www-data. This host identifier and account constitute the known entities used for subsequent retrieval.
[0026] S1. Event Modeling and Candidate Recall S1-1 retrieves multi-source logs and alarms related to the security event to be analyzed, maps the fields of each data source to a unified structure and attaches original record references to obtain a unified event node set.
[0027] The field mapping module first assigns an event number to each input record and maps synonymous fields from different data sources to occurrence time, data source identifier, data view, subject entity, object entity, action type, session identifier, transaction identifier, and original record reference. Host audit logs map process number, parent process number, session number, executing account, working directory, and command; authentication logs map authentication type, authentication result, operating account, target account, and session number; web access logs map client address, request method, resource path, service identifier, and response status; and network flow records map network flow number, source address, source port, destination address, destination port, and protocol.
[0028] Alarm events are represented as follows: A=<alert_id,time,type,source_entity,target_entity,host,account,resource,severity,detector_id> Wherein, alert_id is the alarm number, time is the alarm occurrence time, type is the alarm type, source_entity is the source entity marked by the alarm, target_entity is the target entity marked by the alarm, host is the host identifier to which the alarm belongs, account is the account associated with the alarm, resource is the resource involved in the alarm, severity is the alarm level, and detector_id is the detector number that generated the alarm.
[0029] Context log events are represented as follows: E=<event_id,time,source,view,entities,action,session_id,transaction_id,raw_ref> Wherein, event_id is the event number, time is the time the event occurred, source is the log source, view is the data view to which the event belongs, entities is the set of entities involved in the event, action is the event action type, session_id is the session number, transaction_id is the transaction number, and raw_ref is the original record reference.
[0030] The field mapping module only writes fields that can be parsed from the original record. Fields not provided in the original record are left empty. It does not infer entities based on proximity of time or similar addresses.
[0031] The raw record reference (raw_ref) includes the data source identifier, file or table identifier, record location, occurrence time, and raw content summary. The event number is determined based on the scene identifier, data source identifier, record location, occurrence time, and raw content summary, ensuring that the same input record receives the same event number when processed repeatedly.
[0032] In this embodiment, event E1 is an alarm pending analysis; event E2 is a successful su record in the authentication log, recording that www-data switched to the account blord; event E3 is a sudo record, recording that blord executed a command with root as the target account, and includes the terminal, working directory, and command content; event E4 is a USER_CMD record in the host audit log, including the process number, terminal, working directory, and decoded command. E2 to E4 retain their log source, record location, and original content summary, respectively.
[0033] S1-2 Based on a unified event node set, events with shared connection keys are found using explicit connection keys, and candidate contexts are recalled by shared address, shared host, or time proximity to obtain an extended event set.
[0034] Candidate extensions begin with the host identifier and account in E1. The retrieval module first uses session number, request number, network flow number, process number, parent-child process identifier, account switch record, and process-socket mapping to find events with clear connection criteria. For network data, this embodiment uses the flow_id in the Suricata record as the network flow number; the source address, source port, destination address, destination port, and protocol only form the connection key when the records can identify the same connection.
[0035] The retrieval module then recalls other candidate events based on shared addresses, shared hosts, or a retrieval interval centered on the alarm occurrence time. This recall only expands the records to be inspected and does not directly confirm attack relationships. Records with only shared addresses, shared hosts, or within the same retrieval interval can only generate weak contextual relationships without explicit join keys and behavioral rules.
[0036] The extended event set is deduplicated by event number, and each event is recorded as being recalled by a seed entity and a join key. Records recalled by an explicit join key are entered into the corresponding view for relation rule matching; records recalled by shared address, shared host, or time proximity also retain the recall reason for use in generating weak context relations.
[0037] S2, Construction of Layered Evidence Map S2-1 applies the hierarchical relationship generation rule to single event records, event record pairs, and event record sequences in the extended event set to obtain a hierarchical relationship set.
[0038] Relationship generation rules are executable decision conditions that convert one or more event records into directed relationships in a candidate evidence graph. Each rule includes a rule number, applicable data view, input event type, participating decision fields, field value conditions, event connection conditions, output relationship type, semantic level, supporting event record method, and claim boundary. In this embodiment, the rules are embedded in the relationship construction module, eliminating the need for operations personnel to configure rules for each analysis.
[0039] Direct operation layer relationships are denoted as P-layer, representing operations that can be directly verified on a single record; complete transaction layer relationships are denoted as T-layer, representing transactions with explicit join keys and identifiable start and end points; multi-event behavior layer relationships are denoted as B-layer, representing behaviors formed after multiple P-layer or T-layer relationships satisfy subject connection, object connection, and event sequence; weak context layer relationships are denoted as C-layer, representing candidate relationships formed by shared addresses, shared hosts, or temporal proximity. Semantic levels define the permissions of relationships to participate in critical paths and support conclusions, but do not indicate the data source category.
[0040] For a relation rule *r* and its input events, the relation building module sequentially executes field conditions, relation conditions, and lookupable conditions. Field conditions verify the subject, object, action, time, and data source fields specified by the rule; relation conditions verify the join key, session, transaction, and action order; and lookupable conditions confirm that each supporting event has an event number and an original record reference. When all three conditions are met, a P-level, T-level, or B-level relation is generated, and the rule number, participating decision fields, and supporting event number are written. Events that do not satisfy a strong relation rule do not automatically generate a C-level relation; a C-level relation is only generated when events actually share a common address, share a common host, or are close in time.
[0041] Each relation record includes relation_id, source_node_id, target_node_id, relation_type, view, layer, rule_id, support_event_ids, source_fields, time_gap, and claim_boundary. relation_id is the relation ID; source_node_id and target_node_id identify the two ends of the relation, respectively; support_event_ids records the event IDs used to generate the relation; source_fields records the fields used in the decision-making process; time_gap records the time difference between the later and earlier events in a multi-event relation; and claim_boundary records the scope of facts that the relation can directly support.
[0042] The host and authentication views execute the following rules: A P-level account switching relationship is generated when the authentication log contains a successful su record and can resolve the source and target accounts; a P-level authorized command relationship is generated when the sudo record contains the executing account, target account, and COMMAND fields; a P-level command execution relationship is generated when the USER_CMD record contains a resolvable command; a T-level session lifecycle relationship is generated when the PAM session start and end records have the same host, process ID, session ID, account, and program path, and the end record occurs after the start record.
[0043] The host behavior rules connect account switching relationships, authorization command relationships, and command execution relationships in event sequence. This rule requires that the target account of `su` is consistent with the executing account of `sudo`, that the commands, terminals, and working directories in the `sudo` record are consistent with those in the `USER_CMD` record, and that each component relationship can be traced back. When the conditions are met, a B-level permission operation chain relationship is generated, with its supporting event number being the `su`, `sudo`, and `USER_CMD` event numbers that make up the chain. In this embodiment, E2, E3, and E4 satisfy this rule, forming a behavior chain where `www-data` executes root-privilege commands via `blord`.
[0044] In the Web view, when an access record simultaneously contains the client address, request method, resource path, and response status, a Layer T Web transaction relationship is generated. This relationship only represents the request and response status appearing in the record. When the same service and the same client access two different preset risk paths within 60 seconds, and both paths match the path pattern in the rules module, a Layer B path access sequence relationship is generated. This relationship only supports the fact of continuous access and does not directly prove successful exploitation. When only the client address is shared but lacks request identifiers, connection identifiers, or behavioral rules, a Layer C relationship is generated.
[0045] In the network view, when a record contains flow_id, source endpoint, destination endpoint, and protocol, a Layer T network flow relationship is generated. Cross-view connections only use request number, trace number, network flow number, session number, or process-socket mapping that can appear in both views simultaneously. When cross-view records only share IP addresses, a Layer C relationship is generated, and network flow transactions are not directly incorporated into the host behavior chain.
[0046] S2-2 Construct a candidate evidence graph using event nodes in the extended event set as vertices and hierarchical relationship sets as edges.
[0047] like Figure 2 and Figure 3As shown, the candidate evidence graph is organized into host and authentication views, web views, and network views. Each view stores the nodes and relationships that the data source can directly support, and cross-view relationships are connected to the corresponding views through explicit join keys. The relationship direction is determined according to the subject-to-object direction in the records; multi-event relationships are determined according to the order of occurrence of the supporting events.
[0048] The candidate evidence graph retains P-layer, T-layer, B-layer, and C-layer relationships and their supporting records. P-layer, T-layer, and B-layer relationships that satisfy the integrity rules can participate in ordered critical path search; C-layer relationships are used to display candidate connections and continue recall context, but do not participate in path closure independently. When a relationship lacks support_event_ids, source_fields, or original record references, it is marked as an unretrievable relationship and excluded from the core subgraph candidate set.
[0049] In this embodiment, E2 generates a P-layer account switching relationship R1 from www-data to blord, E3 generates a P-layer relationship R2 from blord to root authorization commands, and E4 generates a P-layer command execution relationship R3 from process instances to command nodes; E2, E3, and E4 jointly generate a B-layer permission operation chain relationship R4. Network flows that share an address with this host but do not have an explicit connection key are retained as C-layer relationships.
[0050] S3, Minimum Sufficient Evidence Subgraph and Evidence Generation S3-1 Obtain event seeds from known entities in the extended event set, search for ordered critical paths in the candidate evidence graph, and remove unnecessary relations under sufficiency constraints to obtain the minimum sufficient evidence subgraph. The minimum sufficient evidence subgraph extraction process is as follows: Figure 4 As shown.
[0051] Event seeds are derived from known entities in the metadata of the event to be analyzed. This embodiment uses the host identifier and the www-data account in E1 as event seeds. The rule template specifies the key behavior types that must be covered according to the type of event to be analyzed; in this embodiment, the privilege escalation template specifies three key behavior types: account switching, authorized execution, and sensitive resource operations. Sensitive resource operations are determined by the / etc / shadow or / root resource explicitly appearing in the command content, without inferring resource access based on account permissions. / etc / shadow is a sensitive system file in the Linux operating system that stores user account names and corresponding password hashes; only root users have read permissions, making it a core sensitive resource of the operating system.
[0052] Path searching starts from the event seed and extends only along the P-level, T-level, and B-level relationships that satisfy the complete behavioral rules. Key behaviors are connected according to the relationship direction and the order of supporting events until the affected object is reached. Evidence closure means that the obtained path starts from the event seed, reaches the affected object through the key behaviors specified in the template, and each key relationship has a supporting event number and an original record reference. A path consisting entirely of C-level relationships does not satisfy evidence closure.
[0053] For each candidate subgraph H, the key behavior coverage rate is the ratio of the number of key behavior types covered by H to the number of key behavior types specified in the template; the relationship recall rate is the proportion of nodes and relationships that simultaneously have supporting event numbers and original record references to the total number of nodes and relationships in H; the redundant relationship ratio is the proportion of duplicate transaction relationships and relationships whose deletion does not affect evidence closure to the total number of relationships in H. Candidate subgraph selection considers key behavior coverage rate, relationship recall rate, redundant relationship ratio, and number of relationships simultaneously; when target values are the same, the results are selected according to the order of relationship numbers.
[0054] The specific extraction process is as follows: First, exclude C-level relationships and non-retrievable relationships; then search for closed paths in the covering rule template; subsequently, add the P-level or T-level supporting relationships that the path directly depends on; finally, attempt to delete the remaining relationships one by one according to their relationship numbers. If deleting a relationship still covers the key behaviors specified in the template, the key relationships are still retrievable, and the path remains closed, confirm the deletion of the relationship; restore the relationship if any constraint no longer holds. After completing one round of checks, continue performing the same checks on the remaining relationships until no relationship can be deleted while maintaining sufficiency.
[0055] The sufficiency requirement for the minimum sufficient evidence subgraph is that it covers the key behaviors specified in the template, has a closed path from the event seed to the affected object, and the key relationships in the path can be traced back. The minimum requirement is that after deleting any unnecessary relationship, the above conditions are no longer met simultaneously. If there is no path in the candidate evidence graph that satisfies the sufficiency requirement, a closed subgraph is not generated, the uncovered key behaviors and missing connections are recorded, and the evaluation status is set to insufficient evidence.
[0056] In this embodiment, the ordered critical path is as follows: the www-data account is switched to blord via R1, blord obtains root command execution privileges via R2, and the related process executes the command containing the / etc / shadow resource via R3; R4 records the permission operation chain consisting of three supporting events. Shared host records, shared address network flows, and records within the same session that do not participate in this path are excluded during the edge-by-edge deletion process. The resulting subgraph retains R1 to R4 and their directly supporting events.
[0057] S3-2 Generates evidence credentials based on the minimum sufficient evidence subgraph and extracts the set of evidence numbers that the judgment module is allowed to cite from the evidence credentials.
[0058] The evidence generation module assigns unique numbers to nodes and relationships in the subgraph for this assessment. Node numbers use N followed by an incrementing sequence number, and relationship numbers use R followed by an incrementing sequence number. Evidence includes subgraph number, node number, node type, relationship number, relationship semantic level, rule number, ordered critical path, supporting event number, fields involved in the judgment, original record references, and missing evidence.
[0059] The set of allowed evidence numbers consists of node numbers and relation numbers in the minimum sufficient evidence subgraph, excluding deleted relations, C-level candidate relations, and original records not included in the subgraph. Original record fragments are mapped to evidence numbers through supporting event numbers, ensuring that each allowed relation can be located to the log record that generated that relation.
[0060] This embodiment assigns node numbers to www-data, blord, root, process instances, commands, and sensitive resources, and assigns relationship numbers to account switching, authorization commands, command execution, and permission operation chains. The allowed reference set is a finite set of the node numbers and relationship numbers. The original record fragments of E2, E3, and E4 are appended to the corresponding relationships according to support_event_ids.
[0061] S4. Evidence Constraint Assessment and Verification The overall process is as follows Figure 5 As shown: S4-1 Input the minimum sufficient evidence subgraph, the original record fragments it references, and the set of evidence numbers that can be referenced into the analysis module to generate a structured analysis result.
[0062] The input for analysis consists of four parts: the event field to be analyzed, the minimum sufficient evidence subgraph, the original record fragments referenced by the subgraph, and the set of evidence numbers that can be referenced. Nodes are serialized according to node numbers, and relations are serialized according to the field order of "relation number, source node number, relation type, target node number, semantic level, rule number, and supporting event number". Ordered critical paths are listed separately according to the relation number order.
[0063] The analysis module employs a general language model. Input prompts explicitly state: analysis reasons can only cite evidence numbers from the allowed set; C-level relationships cannot be the sole basis for attack conclusions; and insufficient evidence must be output when a closed path cannot be formed from allowed evidence. The range of evidence numbers is implemented through an evidence number whitelist, structured output mode, JSON Schema constraints, syntax-guided decoding, output tag masks, or probability biases. Specific deployments can choose one of these implementations or use a combination of them.
[0064] The structured analysis results include the fields decision, confidence, reason, evidence_ids, and missing_evidence. The decision field represents the analysis conclusion, with values including attack, benign, and insufficient_evidence. Attack indicates the event is judged as an attack, benign indicates it is judged as a normal event, and insufficient_evidence indicates insufficient existing evidence to make a judgment. Confidence represents the confidence level of the analysis conclusion. Reason represents the rationale behind the conclusion. Evidence_ids represents the evidence numbers cited in the rationale, requiring only the node and relation numbers directly supporting the reason. Missing_evidence represents missing evidence, indicating behaviors or connections still missing after completing the closed path. The analysis module does not directly accept relationships that have been deleted from the candidate evidence graph.
[0065] In this embodiment, the analysis module generates attack conclusions based on R1 to R4. The reason explains that after www-data is switched to blord, it executes root privilege commands and applies them to / etc / shadow. The evidence_ids is filled with the node number and relationship number directly used in this conclusion, and missing_evidence is empty.
[0066] S4-2 uses the set of allowed evidence numbers as the allowed set and the ordered critical path as the closure criterion to perform independent evidence verification on the structured analysis results and output the final analysis conclusion.
[0067] The independent evidence verification module is located at the output of the judgment module and does not share the intermediate reasoning content of the judgment module. The verification module first calculates the difference between evidence_ids and the allowed reference set. If the difference set is not empty, the number in the difference set is marked as an invalid reference. Then, it reads the semantic level, rule number, supporting event number, and claim boundary of the relationship corresponding to the valid number and verifies whether the relationship can support the corresponding behavior in the reason. Finally, it checks the starting point, key behavior, affected object, relationship direction, and original record reference according to the ordered critical path.
[0068] The validation passes when all evidence_ids belong to the allowed reference set, the key behaviors in the judgment reason are all supported by relationships, and the ordered critical path is closed. The structured judgment result is output as the final judgment conclusion. If any item fails, the validation module outputs invalid references and missing evidence. This validation module can consist of deterministic set checks, graph path checks, and semantic correspondence checks, and does not rely on the judgment module to confirm the correctness of its output.
[0069] If the first verification fails, invalid references, missing evidence, and the original set of allowed references are returned to the decision module for a second inference. The second inference still only accepts the original minimum sufficient evidence subgraph, the original record fragment, and the original set of allowed references, without adding relationships to the candidate evidence graph. The result of the second inference is subjected to the same verification process again; if it still fails, the decision is corrected to insufficient evidence, and the verified evidence, unclosed links, and corresponding original record references are output.
[0070] In this embodiment, the independent evidence verification module confirms that the numbers cited in the analysis results are all within the allowed set, and R1 to R4 have supporting records of E2, E3 and E4 respectively. The ordered critical path starts from www-data, reaches / etc / shadow through blord and root privilege commands, and thus outputs the attack conclusion. The critical path, valid evidence and original record citations are then delivered to the analyst for review.
[0071] Based on the above method, this embodiment also provides a security event analysis system based on a minimum sufficient evidence subgraph. The system includes a unified event representation module, a candidate event expansion module, a hierarchical relationship construction module, a candidate evidence graph generation module, a minimum sufficient evidence subgraph extraction module, an evidence credential generation module, an analysis module, and an independent evidence verification module.
[0072] The unified event representation module performs field mapping and generates original record references; the candidate event extension module recalls events using explicit join keys and weak context conditions; the hierarchical relationship construction module executes relationship rules and generates P-layer, T-layer, B-layer, and C-layer relationships; the candidate evidence graph generation module organizes nodes and relationships according to the data view; the minimum sufficient evidence subgraph extraction module searches for closed paths and performs edge-by-edge deletion; the evidence credential generation module assigns evidence numbers and forms a set of allowed references; the analysis module generates structured results; and the independent evidence verification module checks the number attribution, behavioral support, and path closure.
[0073] Each module can be deployed on the same computing device, or on multiple computing devices capable of transmitting unified event records, relationship records, evidence subgraphs, evidence credentials, and judgment records. Graph data can be stored in a graph database, or in a file or relational database capable of storing nodes, edges, and references to original records. These storage methods do not alter the hierarchical relationship permissions, the extraction of the minimum sufficient evidence subgraph, or the evidence constraint judgment process.
[0074] The experimental data in this embodiment comes from the AIT Alert Dataset and AIT Log Dataset V2.0 datasets. The AIT Alert Dataset provides AMiner alerts and attack time interval annotations, while the AIT Log Dataset V2.0 provides host audit logs, authentication logs, web access logs, DNS logs, and Suricata network logs corresponding to each scenario.
[0075] In the experimental data, the candidate evidence graph contained 1,546,662 relations, and the minimum sufficient evidence subgraph retained 65 relations; 100 correctly matched samples did not form erroneous closed subgraphs. These results indicate that hierarchical relation permissions and subgraph extraction can compress candidate relations and prevent weak context from forming closed conclusions on its own.
[0076] Experimental results are as follows Figure 6 As shown, in the comparative analysis of 45 cases, the accuracy rate of the minimum sufficient evidence subgraph combined with the original records was 95.2%, with an F1 score of 0.929. The shared entity relationship graph, which is a comparison graph formed by establishing relationships between events containing the same address, host, account, or resource within the same candidate event set using traditional methods, achieved an accuracy rate of 66.7% when combined with the original records. The average input size of the minimum sufficient evidence subgraph was 8.8% of that of the shared entity relationship graph. In the 42 jointly completed cases, 12 cases were changed from incorrect to correct, and no cases were changed from correct to incorrect.
[0077] In addition, in 17 cases comparing tiling input and graph input containing the same original events, the accuracy of graph input changed from 76.47% to 82.35%, the attack recall rate changed from 55.56% to 66.67%, the accuracy of both attack groups was 100%, and the false positive rate of normal input was 0.
[0078] In this embodiment, graph data is stored in a graph database, and each module is deployed on the same computing device. The modules transmit unified event records, relationship records, minimum sufficient evidence subgraphs, evidence credentials, and analysis records.
[0079] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
[0080] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A security event assessment method based on minimum sufficient evidence subgraphs, characterized in that, The process includes the following: S1. Obtain multi-source logs and alarms involved in the security event to be analyzed, and represent them as a unified set of event nodes with additional original record references. Use explicit join keys to expand the context and recall candidate events to obtain an expanded event set. S2. Execute the hierarchical relationship generation rule on the extended event set to generate a hierarchical relationship set containing the direct operation layer, the complete transaction layer, the multi-event behavior layer and the weak context layer. Construct a candidate evidence graph with event nodes as vertices and hierarchical relationships as edges. S3. Extract the minimum sufficient evidence subgraph from the candidate evidence graph that covers key behaviors, has traceable relationships, and has a closed path. Assign unique numbers to the nodes and relationships in the minimum sufficient evidence subgraph, generate evidence credentials, and form a set of evidence numbers that can be cited. The minimum sufficient evidence subgraph extraction process is as follows: Determine the ordered critical path corresponding to the event to be assessed. The ordered critical path starts from the event seed, passes through the key behavior, and reaches the affected object. The key relationships on the path belong to one or more of the following types: direct operation layer relationship, complete transaction layer relationship, and multi-event behavior layer relationship. Each key relationship records a supporting event number and an original record reference. Paths composed entirely of weak context layer relationships do not satisfy the path closure condition. Under the condition of maintaining key behavior coverage, traceable relationships, and closed evidence, delete unnecessary relationships from the candidate evidence graph to obtain the minimum sufficient evidence subgraph. Assign a unique number to each node and each relationship in the minimum sufficient evidence subgraph, generate evidence credentials including subgraph number, node number, relationship number, ordered critical path, supporting event number, and original record reference, and extract node number and relationship number from the evidence credentials to form the set of evidence numbers that can be cited. S4. Input the minimum sufficient evidence subgraph, the original records it references, and the set of evidence numbers that can be referenced into the constraint judgment module. Use the general large model to perform constraint judgment and generate judgment results. Perform independent evidence verification on the judgment results. If the verification passes, output the judgment results directly. If the verification fails, return to the constraint judgment module for secondary reasoning. If it still fails, output insufficient evidence.
2. The security event assessment method based on the minimum sufficient evidence subgraph as described in claim 1, characterized in that: In step S1, alarm event nodes are extracted from alarms to be analyzed, context logs are parsed from multi-source logs and normalized into a unified event structure, context log event nodes are constructed, and original record references including log source, record location, occurrence time, and original content summary are attached to each event node to obtain the unified event node set; the explicit connection key includes session number, request number, network flow number, process number, parent and child process identifier, account switching record, network 5-tuple, and process-socket mapping.
3. The security event assessment method based on the minimum sufficient evidence subgraph as described in claim 1, characterized in that: In S2, the hierarchical relationship generation rule is as follows: For a single event record, event record pair, and event record sequence in the extended event set, first determine whether the value requirements of the matching relationship rule are met in the time, subject, object, action, and data source fields; then determine whether the association requirements of the matching relationship rule are met in the connection key, session, transaction, and behavior sequence; finally, determine whether the event pair has a supporting event number and original record reference. When all three judgments pass, the rule number of the matching relationship rule and the fields involved in the judgment are written into the relationship record and the relationship is generated according to the corresponding semantic level. If any judgment fails, only a weak context layer relationship is generated.
4. The security event assessment method based on the minimum sufficient evidence subgraph as described in claim 1, characterized in that: In S2, the direct operation layer relationship represents an operation that can be directly verified for a single record; the complete transaction layer relationship represents a closed transaction consisting of a request and response, a connection start point and a connection end point, and an authentication operation and authentication result. Multi-event behavior layer relationships represent a sequence of behaviors formed by connecting direct operation layer relationships and complete transaction layer relationships; weak context layer relationships represent candidate connections formed by shared addresses, shared hosts, or time proximity.
5. The security event assessment method based on the minimum sufficient evidence subgraph as described in claim 1, characterized in that: The construction of the candidate evidence graph in S2 is specifically as follows: the event nodes in the extended event set are used as vertices, the relationships in the hierarchical relationship set are used as edges, and the graph is organized according to the data source view to form the candidate evidence graph; the relationships within the same data source view form the subgraph within the view, and the relationships across data source views are used to connect the subgraphs within the view.
6. The security event assessment method based on the minimum sufficient evidence subgraph as described in claim 5, characterized in that: The data source views include four categories: host and authentication views generate relationships based on process ID, parent and child process IDs, session ID, account switching records, and command execution records; web views close the same request transaction using request ID, connection ID, client address, service ID, and resource path; network views close connections using network flow ID or network 5-tuple, and close domain name resolution requests and responses using transaction IDs. Cross-view associations use request numbers, trace numbers, network flow numbers, session numbers, or process-socket mappings that can appear in two views simultaneously.
7. The security event assessment method based on the minimum sufficient evidence subgraph as described in claim 1, characterized in that: The minimum sufficient evidence subgraph is selected from the candidate evidence graphs according to the following objective function: ; ; ; ; Where G is the candidate evidence graph, H is the extracted critical path subgraph, and G* is the minimum sufficient evidence subgraph; Coverage(H) is the critical behavior coverage rate, representing the proportion of the number of critical behavior types covered by subgraph H to the number of critical behavior types specified in the rule template; Trace(H) is the relationship traceability rate, representing the proportion of nodes and relationships in subgraph H that have original record references or supporting event references to the total number of nodes and relationships in subgraph H; Redundancy(H) is the redundant relationship ratio, representing the proportion of duplicate transaction relationships and relationships in subgraph H that, after deletion, do not affect the closure of evidence to the total number of relationships in subgraph H; n cover (H) represents the number of key behavior types covered by subgraph H, n required n represents the number of key behavior types specified in the rule template corresponding to the event to be analyzed. trace (H) represents the number of nodes and relationships in subgraph H that have original record references or supporting event references, n. redundant (H) represents the number of duplicate transaction relations in subgraph H and the number of relations that do not affect the closure of evidence after deletion. |V(H)| and |E(H)| represent the number of nodes and the number of relations in subgraph H, respectively. λ is the penalty coefficient for the number of relations. For candidate subgraphs with the same target value, the subgraph with the relation number sorted first is selected.
8. The security event assessment method based on the minimum sufficient evidence subgraph as described in claim 1, characterized in that: The data processing flow of the constraint analysis module is as follows: The minimum sufficient evidence subgraph, the original record fragments it references, and the set of allowed evidence numbers are input into the constraint judgment module. The general large language model is used to generate the judgment result. By using one or more of the following methods, such as the evidence number whitelist in the prompt words, structured output mode, JSON Schema constraints, syntax-guided decoding, output tag mask, and probability bias, the range of evidence numbers referenced in the judgment process is restricted, and a structured judgment result including the judgment conclusion, confidence level, reasoning, referenced evidence numbers, and missing evidence is generated.
9. The security event assessment method based on the minimum sufficient evidence subgraph as described in claim 1, characterized in that: The specific process for performing independent evidence verification on the assessment results is as follows: Each cited evidence number is checked to ensure it belongs to the set of allowed evidence numbers. The semantic level, rule number, and supporting event number of each number's corresponding relationship are read to determine whether the cited relationship supports the reasoning. It is also checked whether the ordered critical path starts from the event seed, reaches the affected object through the critical behavior, and whether the critical relationships on the path all have supporting events and original record references. If the verification passes, the structured judgment result is output as the final judgment conclusion. If the verification fails, the invalid references and missing evidence are returned to the constraint judgment module for secondary reasoning. The evidence numbers cited in the secondary reasoning are still limited to the set of allowed evidence numbers. After the secondary reasoning, independent evidence verification is performed again. If it still fails, the conclusion is corrected to insufficient evidence and output, along with the critical path, valid evidence, and original record references.
