함정 폐쇄망 조건을 고려한 LLM 시스템 구축과 운용 시나리오

옥지원¹^{, *} ; 이정식² ; 한인성²

1국방과학연구소 제3기술연구원 연구원
2국방과학연구소 제3기술연구원 책임연구원

Design and Operational Scenarios of LLM Systems for Closed-network Naval Platforms

Jiwon Ock¹^{, *} ; Jungsik Lee² ; Insung Han²

11Researcher, 3rd R&D Institute, Agency for Defense Development
2Principal researcher, 3rd R&D Institute, Agency for Defense Development

Correspondence to: ^*Jiwon Ock Tel: +82-2-3400-2585 Fax: +82-2-403-3512 E-mail: jwock@add.re.kr

초록

본 논문에서는 해군 함정과 같은 폐쇄망 환경에서 운용 가능한 LLM 시스템 구축 방안을 제안한다. 제한된 연산 자원과 네트워크 단절 조건을 고려하여, MCP(model context protocol)와 RAG(retrieval-augmented generation) 기술을 통합한 아키텍처를 설계하고, 실시간 문서 기반 질의응답이 가능한 프로토타입 시스템을 구현하였다. 또한, 해군 내 운용 시나리오를 기반으로 LLM의 적용 가능성을 확인하였다. 오픈소스 모델 활용을 위한 GPU 자원 분석으로 제한된 하드웨어를 탑재한 함정에서 LLM을 도입하게 되었을 때 참고지표로 활용할 수 있다.

Abstract

This study proposes an LLM system for closed network environments like Navy ships, addressing limited hardware and no-network conditions. By combining MCP and RAG techniques, a prototype for real-time, document-based Q&A was developed and validated through a Navy operational scenario. Additionally, by analyzing GPU resource requirements of open-source LLMs, we provide reference benchmarks for implementing such systems on ships with constrained hardware capabilities.

Keywords:

Closed-Network of Battleship, Large Language Model, Model Context Protocol, Retrieval-augmented Generation, Navy’s Operational Scenarios

키워드:

함정 폐쇄망, 거대 언어모델, 모델 컨텍스트 프로토콜, 검색 증강 생성, 해군 운용 시나리오

Acknowledgments

이 논문은 2025년도 한국해군과학기술학회 하계학술대회 발표 논문임.

References

Kapiamba, S., Fouad, H., and Moskowitz, I. S., “Responsible Integration of Large Language Models (LLMs) in Navy Operational Plan Generation,” Proceedings of the AAAI Symposium Series, Vol. 3, No. 1, 2024, pp. 50-53. [https://doi.org/10.1609/aaaiss.v3i1.31179]
Lee, S. U., Lee, S. H., Yu, H. C., Park, S. S., and Kim, U. H., “Analyzing and Addressing Challenges of LLM Applications in Defense,” Journal of Defense and Security, Vol. 6, No. 1, 2024, pp. 302-328.
Jim Halsell, “Generative Large Language Models of Navy Crews,” U.S. Naval Institute Proceeding, Vol. 151, No. 468, 2025.
Xinyi, H., Yanjie, Z., Shenao, W., Haoyu, W., “Model Context Protocol (MCP): Landscape, Security Threats, and Future Research Directions,” arXiv preprint arXiv:2503.23278, , 2025.
Lewis, Patrick, et al. “Retrieval-augmented generation for knowledge-intensive nlp tasks,” Advances in neural information processing systems, Vol. 33, 2020, pp. 9459-9474.
Hugging Face, https://huggingface.co/, , Accessed June 9, 2025.
L. Tuggener, P. Sager, Y. Taoudi-Benchekroun, B. F. Grewe and T. Stadelmann, “So You Want Your Private LLM at Home?: A Survey and Benchmark of Methods for Efficient GPTs,” 2024 11th IEEE Swiss Conference on Data Science (SDS), Zurich, Switzerland, 2024, pp. 205-212. [https://doi.org/10.1109/SDS60720.2024.00036]
Meta, “The llama 3 hred of models,” arXiv preprint arXiv:2407.21783, , 2024.
Google Team, “Gemma3 technical report,” arXiv prepring arXiv:2503.19786, , 2025.
Mistral Ai, “Mistral 7B,” arXiv preprint arXiv:2310.06825, , 2023.
Qwen Team, “Qwen technical report,” arXiv preprint arXiv:2309.16609, , 2023.