Communication-efficient Federated Transfer Learning for real-time intrusion detection in agricultural IoT systems

Abstract

The rapid deployment of Agricultural Internet of Things (Ag-IoT) networks has introduced significant security challenges, mainly due to their limited computational resources and operation in remote environments. These constraints, combined with the critical nature of agricultural data and operations, make Ag-IoT systems particularly vulnerable to cyber threats, thus requiring robust, efficient security mechanisms. In this context, Intrusion Detection Systems (IDS) play a crucial role in monitoring network activities and detecting malicious behavior. However, conventional IDS solutions often suffer from high computational overhead, increased latency, and substantial communication costs. Although Federated Learning (FL) enhances data privacy by enabling distributed model training, its efficiency remains constrained in resource-limited Ag-IoT scenarios. To overcome these limitations, we propose a novel IDS based on Federated Transfer Learning (FTL) for multiclass network intrusion classification, which integrates FL with lightweight pretrained image models. The key idea of our approach is to transform network traffic data into grayscale images, allowing the reuse of efficient computer vision models. This transformation reduces data complexity, which is expected to support low-resource training and inference on edge devices. The FTL framework ensures that only model updates are exchanged, minimizing communication overhead while preserving data privacy. Experimental results, validated over three independent runs (Mean ± SD), demonstrate that our approach achieves 99.96% ± 0.00% accuracy on the Edge-IIoTset and 99.96% ± 0.01% accuracy on the Farm-Flow dataset. These findings highlight the strong theoretical potential of our FTL-based IDS as a promising candidate to enable intrusion detection in real-time, resource-efficient, and privacy-preserving Ag-IoT networks.