Adaptive Zero-Trust Microsegmentation via Graph-Aware Learning for Wireless Networks

Abstract

Large-scale wireless networks enable dense multihop connectivity but also expose new pathways for lateral attack propagation. This paper develops a theory-guided adaptive microsegmentation framework that contains attack spread while preserving network connectivity. Grounding the framework in the zero-trust principle of continuous verification, we treat each isolation decision as a per-epoch authorization outcome derived from observed device posture, rather than as a static security label. Modeling the network as a random geometric graph, we formulate microsegmentation as a constrained graph-control problem that suppresses epidemic percolation under limited security budgets. The formulation captures the security-performance tradeoff through spectral radius reduction, giant component suppression, and residual communication utility. To solve this high-dimensional problem, we propose a graph-attention-based reinforcement learning approach that learns topology-aware isolation policies. Simulation results demonstrate that the proposed method significantly reduces large-scale attack propagation while maintaining higher connectivity compared to heuristic and nongraph learning baselines.