Fractional Groupers Brown-Bear Optimization based Spiking Channel Attention-based Convolutional Recurrent Neural Network for Intrusion Detection in Heterogeneous Federated IoT Networks

In Internet of Things (IoT) networks, detecting intrusions is vital to safeguard interconnected devices from unauthorized access. Poor scalability and privacy concerns when aggregating sensitive information centrally are the issues faced by preceding approaches. Therefore, integrating Federated Learning (FL) is essential, as it enables collaborative model training across distributed IoT nodes without sharing raw data. In this work, the proposed Fractional Groupers Brown-Bear Optimization-based Spiking Channel Attention-based Convolutional Recurrent Neural Network (FGBO_SCACRN-Net) is developed for intrusion detection in heterogeneous IoT environments comprising resource-constrained devices and resource-rich nodes. Within the FL framework, each IoT node trains a local model using its own data and transmits only updated parameters to the central server. For practical deployment, computationally intensive training tasks are primarily executed on gateway or edge nodes, while low-power sensing devices focus on data collection and communication with nearby aggregators. In the training model, intrusion detection is done by SCACRN-Net, which is tuned by Fractional Groupers Brown-Bear Optimization (FGBO). The FGBO is designed by integrating the Fractional Concept (FC) and Groupers Brown-Bear Optimization Algorithm (GBOA). Here, GBOA combines Groupers Moray Eels Optimization (GMEO) and Brown-Bear Optimization Algorithm (BOA). The SCACRN-Net integrates Spiking Neural Network (SPSNet) and Channel Attention-based Convolutional Recurrent Neural Network (CACRN-Net). Finally, an Explainable Artificial Intelligence (XAI) method, like Shapley Additive Explanations (SHAP), is applied to trained models. Moreover, local updation is done at the server using the weighted average method. The efficacy of FGBO_SCACRN-Net is examined by metrics, like accuracy, sensitivity and specificity that attained superior values of 95.78%, 96.21% and 96.00%.