Local LLM-Based Cyber Incident Analysis in Air-Gapped Networks via Teacher–Student Knowledge Distillation and Agentic Orchestration
Recent cyber incidents have become increasingly sophisticated through Living-off-the-Land (LotL) techniques that exploit legitimate behavior and multi-stage attacks. This requires advanced reasoning capabilities to discern the attack contexts within fragmented large-scale logs. However, closed network environments with physical network separation (air-gapped), such as national critical infrastructures, restrict the use of high-performance cloud large language models (LLMs), thereby limiting the adoption of cutting-edge artificial intelligence (AI)-based analysis technologies. To overcome these constraints, this study proposes a Local LLM-based intrusion analysis framework that operates independently within closed networks. The proposed framework combines (i) an Offline Knowledge Distillation technique that transfers the analytical reasoning process of external high-performance models to the Local LLM after a security review, and (ii) an AI agent orchestration structure that controls the analysis procedure step-by-step and suppresses hallucinations. Experiments and validation using a public dataset (Atomic Red Team) demonstrated that the proposed model achieved a consistently higher detection accuracy (88.4%) and MITRE Adversarial Tactics, Techniques, and Common Knowledge mapping performance (0.91 F1-Score) than existing general-purpose Local LLMs. Furthermore, the proposed model suppressed hallucination rates to 6.2% through an automated verification mechanism and significantly improved analysis efficiency by refining large-scale logs to focus on core events. This study quantitatively demonstrated that AI-based intrusion incident analysis can be automated using a single graphics processing unit server under controlled evaluation conditions. The proposed framework provides a practical prototype for intelligent security monitoring in closed-network environments. However, the operational performance must be validated in real-world deployments.