A federated learning framework for agricultural Internet of Things: sensor data privacy protection and information security

Abstract

The rapid deployment of Internet of Things (IoT) infrastructure in precision agriculture has produced large volumes of distributed sensor data, including soil moisture, ambient temperature, humidity, and crop-health indicators. Although centralized machine-learning pipelines can exploit these data for intelligent decision support, they introduce major concerns related to privacy exposure, transmission overhead, and cyber-security risk. This paper presents FedAgri, a federated learning (FL) framework tailored to agricultural IoT environments. FedAgri enables multiple farm nodes to jointly train a global model for crop-condition prediction without exchanging raw sensor records. To address the heterogeneous and non-independently-and-identically-distributed (non-IID) characteristics of geographically dispersed agricultural data, we introduce a dynamic aggregation mechanism that weights client updates according to local data quality and distribution divergence. In addition, a lightweight differential privacy module is incorporated to provide formal privacy guarantees while preserving model utility. Simulation results on publicly available agricultural datasets show that the proposed framework attains prediction accuracy within 2.3% of a centralized baseline, reduces communication overhead by 68%, and delivers epsilon-differential privacy protection. These findings demonstrate the practical feasibility of privacy-preserving collaborative learning for smart-agriculture applications.