Abstract
<title>Abstract</title> <p> The modernization of smart agriculture increasingly relies on Software-Defined Networking (SDN)-enabled ToT infrastructures to support distributed and data-driven agricultural operations. However, the deployment of heterogeneous and resource-constrained edge devices expands the attack surface while introducing strict requirements related to inference latency, energy efficiency, scalability, and data locality. Existing intrusion detection systems often rely on computationally intensive deep learning models, which may limit their deployment in distributed agricultural ToT environments. <bold>This paper focuses on designing a unified IDS for SDN-based IoT networks in smart agriculture, capable of addressing the extreme demands of an open environment—namely, exposure to external attacks and energy autonomy constraints.</bold> To address these challenges, this paper presents FL-SNNTDS, a federated Spiking Neural Network-based intrusion detection framework. The proposed architecture deploys SNN-based intrusion detection models at local agricultural SDN controllers and uses Federated Learning to collaboratively train a global model without exchanging raw network traffic data. By exploiting sparse spike activity and event-driven computation, the local SNN aims to provide accurate and lightweight intrusion detection with reduced inference latency and operation-level computational cost. The proposed approach was evaluated on the SDN-ToT and Farm-Flow datasets, achieving accuracies of 99.93% and 99.68%, respectively. The federated framework preserved competitive detection performance across distributed clients while maintaining low inference latency and reduced operation-level analytical energy estimates. An external cross-dataset evaluation was additionally conducted using the ASEADOS-SDN-ToT dataset under few-shot adaptation settings. The results demonstrated effective adaptation for dominant traffic categories. Overall, the findings indicate that our proposed <bold>FL-SNNIDS offers optimal trade-off between detection performance, inference latency, energy consumption and operation-level computational complexity.</bold> </p>