Skip to Content
Future InternetFuture Internet
  • Article
  • Open Access

20 November 2026

28 Pages

Fine-Grained ATT&CK Mapping and Attack Graph Construction from Cyber Threat Intelligence via Multi-Instance Reasoning

,
,
and
1
College of Computer Science, Beijing Information Science and Technology University, Beijing 102206, China
2
Institute of Artificial Intelligence Security, Beijing Information Science and Technology University, Beijing 102206, China
3
China Electronic Device System Engineering Corporation, Beijing 100800, China
*
Author to whom correspondence should be addressed.
This article belongs to the Section Cybersecurity

Abstract

Structured modeling of cyber threat intelligence (CTI) is essential for reconstructing attack processes and supporting threat analysis. However, accurately mapping CTI reports to fine-grained MITRE ATT&CK techniques remains challenging because individual events may provide insufficient evidence, while neighboring events can offer complementary context without necessarily sharing identical labels. Existing approaches therefore struggle to exploit contextual dependencies while preserving event-level evidence constraints and independent hierarchical decisions. This paper proposes a context-aware multi-instance reasoning framework for fine-grained ATT&CK mapping and attack graph construction. The framework extracts evidence-grounded atomic attack events from CTI blocks and introduces a constrained multi-instance Tree-of-Thought reasoning mechanism that jointly analyzes related events under shared contextual information and ATT&CK hierarchical constraints while independently determining each event’s tactic, technique, and sub-technique path. An evidence-aware backtracking strategy further prevents unsupported fine-grained predictions by terminating reasoning at a reliable hierarchy level when evidence is insufficient. Experiments on 150 CTI reports show that, compared with AttacKG+ using the same Qwen3-32B backbone, our method improves technique-level F1 from 55.8% to 61.1% and recall from 50.2% to 60.9%. It also achieves 53.2% F1 for sub-technique identification, while the attack-graph analysis assesses temporal consistency under a predefined edge-filtering criterion.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.