Leveraging Knowledge-Driven Large Language Models-Enabled Intent-Based Security Architectures for Human Digital Twin Networks

Ensuring secure and context-aware physical-to-virtual interactions in human digital twin (HDT) networks is crucial because of its demand for adaptive and proactive security mechanisms to support the evolution of virtual environments. Intent-based security design (IBSD) can provide such required proactive alignment of user actions with security policies by embedding intent recognition into the HDT lifecycle, allowing only trusted behavior to propagate while preserving system fidelity. In this paper, we propose a novel knowledge-enabled large language model (kLLM)-driven IBSD architecture that integrates semantic reasoning, continual learning, and contextual understanding to deliver robust dynamic security across distributed HDT networks. The proposed architecture leverages the structured knowledge representation of knowledge graphs (KGs) and the deep semantic capabilities of large language models (LLMs) to enhance intent recognition, threat detection, and adaptive policy enforcement. Such a design also supports the evolution of security models in real-time while preventing malicious updates to virtual replicas in the HDT environment. A case study and comprehensive evaluation demonstrate that the kLLM-enabled IBSD is a promising solution towards efficient intent classification and continual learning adaptability. We also provide key design requirements, enabling techniques, and critical open challenges, including scalability, latency, multimodal intent disambiguation, and KG security. This work represents a significant step toward enabling intelligent, self-adaptive, and resilient security architectures for next-generation distributed digital systems.