ScholarWorks@성균관대학교

초록

Recently, cyberattacks have quickly evolved and transformed from isolated incidents into multiple layers, multi-stage attack chains, posing unprecedented challenges to cybersecurity defenses. Large Language Models (LLMs), with their powerful comprehension and generation capabilities, are emerging as a crucial technological foundation for next-generation intelligent defense systems. However, when confronting complex cyber threats, the LLMs relying solely on understanding and generation remain inadequate. Enhancing their reasoning and decision-making capabilities has thus become a key direction in current cybersecurity research. Focusing on this direction, this paper first outlines evolving demands of cybersecurity and comprehensively reviews existing reasoning techniques. We then integrate GraphRAG into reasoning techniques, proposing an LLM reasoning framework tailored for cybersecurity scenarios. The framework provides models with accurate context and guides their decisions through reasoning techniques. This enhances their reasoning capabilities when confronting complex cybersecurity problems. This enables them to identify threats, reconstruct attack chains, and generate defense strategies, particularly against multi-layered, multi-stage attacks. Subsequently, we validate the effectiveness of our proposed framework through a case study focused on Low Altitude Economy Networks (LAENets) communication. Finally, we summarize potential future research directions in this field.

키워드