A multimodal reinforcement learning-based access control model for power systems under zero-trust architecture

In order to address the challenges of integrating multimodal data and the lack of a comprehensive trust mechanism in power system access control, this paper proposes a novel approach that leverages Zero Trust Architecture (ZTA) principles and multimodal reinforcement learning, named ZTNA-MMRL ( Z ero T rust N etwork A ccess with M ulti M odal R einforcement L earning). Specifically, we design a GraphKAN (Graph Kolmogorov–Arnold Network)-based threat perception mechanism that combines graph neural networks with cross-modal attention to fuse heterogeneous data sources for topology-aware threat detection. We further develop a MoE-DT (Mixture-of-Experts Decision Transformer)-driven access control decision model that adapts to dynamic threat environments through collaborative decision-making among multiple experts. In addition, we propose a five-dimensional trust assessment system that evaluates entities based on authentication reliability, behavioral patterns, contextual rationality, reputation, and risk. Experimental results demonstrate that ZTNA-MMRL achieves 94.7% accuracy, which is 3.4% higher than ZT-Defense, an F1-score of 0.823, an RMSE of 0.142 for trust evaluation, and a response time of 33.1 ms. Additionally, it maintains 86.2% accuracy under data poisoning attacks. These results validate the model's effectiveness in enhancing security, adaptability, and decision-making efficiency in power system access control.