MuCamp2: Generating Validated Cyber Campaign Variants via Constrained LLMs for Group Attribution
Open Access
- 5 June 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 87329-87344
Abstract
As cyberattack operators have progressed to advanced persistent threat (APT) groups, accurate group attribution based on tactic, technique, and procedure (TTP) sequences has gained significant attention. However, publicly available campaign-level TTP sequences remain extremely limited, which restricts both statistical learning and generalization for cyber threat intelligence. Existing augmentation methods, such as synonym replacement (SR), often produce structurally invalid variants because they do not consider cross-technique capability chains. Although large language models (LLMs) offer a promising alternative for generating novel sequences, they alone fail to guarantee sequence integrity. Therefore, the primary limitation of current approaches is not the inability to generate diverse variants, but the tendency to produce a large volume of unusable data. To this end, we propose MuCamp2, a generation-validation architecture for campaign-level TTP augmentation that places LLMs outside the trust boundary. MuCamp2 employs deterministic candidate-pool constraints and a multi-level semantic validation based on the adversary emulation planner (AEP) to improve structural soundness and operational consistency under the available AEP mappings. Experimental results on the Lazarus and MenuPass datasets demonstrate that MuCamp2 completely eliminates tactic violations (0% compared to 26% to 32% for SR) and improves the strict AEP-based validation (L3) pass rate by 5.5 percentage points for Lazarus (10.2% to 15.7%) and 3.0 percentage points for MenuPass (7.5% to 10.5%). Furthermore, downstream attribution achieves a Macro-F1 of 0.987 to 0.990, indicating bounded but consistent usable-yield gains rather than large per-sample quality differences. We further note that L3 should be interpreted as an AEP-based operational proxy: AEP currently covers 197 of 356 ATT&CK v18 techniques (55.3%), and the high L3 pass rate observed for low-coverage sequences reflects reduced validator discrimination rather than higher intrinsic validity.Keywords
Funding Information
- Agency for Defense Development (UG253005TD)
This publication has 27 references indexed in Scilit:
- SynthCTI: LLM-driven synthetic CTI generation to enhance MITRE technique mappingFuture Generation Computer Systems, 2025
- Large Language Models Are Unreliable for Cyber Threat Intelligence2025
- A comprehensive survey of automated Advanced Persistent Threat attribution: Taxonomy, methods, challenges and open research problemsJournal of Information Security and Applications, 2025
- AttacKG+: Boosting attack graph construction with Large Language ModelsComputers & Security, 2024
- LLM-TIKG: Threat intelligence knowledge graph construction utilizing large language modelComputers & Security, 2024
- APT-MMF: An advanced persistent threat actor attribution method based on multimodal and multilevel feature fusionComputers & Security, 2024
- BAN: Predicting APT Attack Based on Bayesian Network With MITRE ATT&CK FrameworkIEEE Access, 2023
- Camp2Vec: Embedding cyber campaign with ATT&CK framework for attack group analysisICT Express, 2023
- IDSGAN: Generative Adversarial Networks for Attack Generation Against Intrusion Detection2022
- Detecting APT Malware Infections Based on Malicious DNS and Traffic AnalysisIEEE Access, 2015