A Neuro-Symbolic Approach to Secure Network Control: Runtime Verification of LLM Decision-Making

Large Language Models (LLMs) are increasingly integrated into network control systems to support anomaly detection, policy synthesis, and adaptive resource management. However, their probabilistic and non-deterministic nature introduces safety risks in security-critical infrastructures such as ۵G and industrial networks. To address this challenge, we propose a neuro-symbolic runtime verification framework that enforces formal safety specifications over LLM-generated control actions. The framework combines neural inference for telemetry interpretation with a symbolic verification layer that encodes network invariants using temporal logic constraints. An abstraction module translates LLM outputs into verifiable predicates, which are then validated by a real-time monitoring engine prior to policy enforcement. We further introduce an adaptive feedback mechanism that refines unsafe decisions through symbolic counterexample analysis. We formally characterize the safety guarantees of the proposed architecture under bounded-latency assumptions and analyze its computational overhead. Experimental evaluation on simulated ۵G slicing and IoT orchestration scenarios demonstrates improved robustness compared to LLM-only baselines, reducing policy violation rates while maintaining low control-plane latency. The resulting verifiable execution traces enhance auditability and trustworthiness in autonomous network management.

کلیدواژه ها:

Neuro-Symbolic AI ، Runtime Verification ، LLMs ، Network Security ، Formal Methods