LLM-assisted federated Bayesian dynamic graph attention network for cloud-edge industrial IoT

Abstract

With the rapid development of cloud-edge industrial Internet of Things, massive sensor and device data are continuously generated at distributed edge clients, while cloud-hosted large language models provide new opportunities for collaborative intelligence and high-level decision support. However, directly uploading industrial data to the cloud may cause privacy leakage, and deploying large models on resource-constrained edge devices is impractical due to high computational and communication costs. To address these challenges, this paper proposes an LLM-assisted federated Bayesian dynamic graph attention network for privacy-preserving graph learning in industrial IoT. In the proposed framework, the Bayesian dynamic graph attention network serves as a lightweight edge-side small model to capture complex relationships among sensors and devices, while the cloud-side LLM acts as a collaborative coordinator that analyzes non-sensitive training statistics and provides high-level guidance for federated optimization. The coordination mechanism is model-agnostic and does not rely on a specific LLM architecture, allowing general-purpose instruction-following LLMs to be integrated as plug-and-play cloud coordinators. Bayesian inference is introduced to model parameter uncertainty through prior and posterior distributions, and a dynamic attention mechanism is used to adaptively learn heterogeneous device relationships. In addition, differential privacy is incorporated into the transmission of model updates to protect sensitive industrial information. Experimental results on multiple graph datasets demonstrate that the proposed method improves federated graph learning performance while maintaining privacy and supporting cloud-edge collaborative intelligence.