Secure Federated Learning in Smart Agriculture: A Survey of Poisoning Attacks, Defences and Challenges

Abstract

Smart agriculture increasingly relies on Federated Learning (FL) to train shared models for crop classification, irrigation scheduling, disease detection, and rainfall prediction across many farms without centralizing sensitive operational data. This decentralization mitigates the privacy and bandwidth costs of traditional cloud-centric machine learning, but FL is well documented to remain vulnerable to poisoning, backdoor, and inference attacks, and agriculture adds threats that are distinctive to the domain, since physical sensors and weather stations are often unattended, easily accessed, and cheaper to tamper with than a data center. This survey organizes the growing but still fragmented literature at the intersection of FL security and precision agriculture, restricting itself throughout to sources published in 2021 or later and independently verified against a real, DOI-registered publication. We review FL fundamentals, catalogue the datasets and applications used in agricultural FL research, present a taxonomy of sensor-level, client-side, communication-side, and server-side attacks, and map robust-aggregation, anomaly-detection, privacy-preserving, and blockchain-audit defences to the attacks they counter. We compare unsecured and secured FL methods on agricultural tasks and identify concrete research gaps: the absence of standardized security benchmarks, limited multi-modal robustness studies, reproducibility limitations, and an under-studied physical-sensor threat model. We close with a research agenda spanning benchmark design, layered defence architectures, multi-modal defences, regulatory-grade audit infrastructure, and edge deployment.