Federated Learning (FL) has emerged as a privacy-preserving alternative to centralized machine learning, enabling collaborative model training without sharing raw data. This is particularly valuable in smart agriculture, where farmers generate sensitive, distributed data that must remain localized due to privacy, legal, and competitive constraints. However, FL introduces new security vulnerabilities stemming from its decentralized nature. This paper investigates these vulnerabilities through a case study involving real-world microclimate time series data from four agricultural greenhouses. We evaluate the effects of Data Poisoning, Model Update Poisoning, and Backdoor Attacks strategies and implement multiple defense mechanisms, including Robust Aggregation, Differential Privacy, and Sanitization. Experimental results demonstrate that while individual defenses can mitigate specific attacks, a comprehensive, multi-layered defense strategy offers superior resilience. We propose a Final Defense framework that combines k-NN-based data sanitization, Trust-Krum aggregation, adaptive clipping, and differential privacy. This configuration consistently achieves the highest F1-Scores across all attack scenarios, enhancing both robustness and convergence efficiency. Our findings confirm the necessity of integrated defenses for secure, scalable deployment of FL in adversarial environments such as smart agriculture.

DOI: https://doi.org/10.1109/GCWkshps68340.2025.11591090