A Federated Learning Framework for IoT Anomaly Detection with Dynamic Client Selection

The proliferation of Internet of Things (IoT) devices has escalated security concerns, making efficient anomaly detection critical. Centralized learning approaches, which require data aggregation, are often infeasible due to privacy constraints, bandwidth limitations, and the distributed nature of IoT ecosystems. This paper proposes a comprehensive Federated Learning (FL) framework designed specifically for distributed anomaly detection in IoT networks. The suggested system incorporates a novel Dynamic Client Selection (DCS) algorithm that prioritizes clients based on data utility and resource capability, moving beyond random selection. We employ a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) hybrid model for spatio-temporal feature extraction from device traffic data. The framework is implemented using TensorFlow Federated (TFF) and evaluated on the N-BaIoT dataset. Experimental results demonstrate that our approach achieves a high detection accuracy of 97.5% and significantly reduces communication rounds by 30% compared to standard FedAvg with random client selection. This resource- efficient and privacy-preserving system is well- suited for real-world deployment in large-scale, heterogeneous IoT environments.