Precise Anomaly Recognition Using Advanced Federated Learning in Resource‐Restricted WSN With Unreliable Connections

ABSTRACT This work presents an advanced federated learning (FL) framework that integrates a CNN–GRU + FedTrust model for precise anomaly recognition in resource‐constrained wireless sensor networks (WSNs) with unreliable connections. Unlike traditional centralized and heterogeneous FL approaches, the proposed system achieves higher accuracy, efficiency, and robustness by combining lightweight CNN–GRU feature extraction with trust‐weighted aggregation through the FedTrust mechanism. The model leverages top‐k gradient sparsification to reduce communication overhead and energy‐aware client selection (EAC) to optimize energy use, ensuring sustainable network performance. Evaluations on real‐world datasets—WUSTL Wireless Sensor Data and Intel Lab Data—demonstrate the model's superiority, achieving 98.4% accuracy, 0.98 F1 score, and faster convergence than Autoencoder‐FL, GNN‐FL, and hierarchical FL methods. It also exhibits enhanced robustness under client dropout (96.1%) and noise (92.7%) conditions, significantly outperforming existing FL techniques. By efficiently capturing both spatial and temporal patterns while maintaining privacy and energy balance, the CNN–GRU + FedTrust framework delivers reliable and scalable anomaly detection across diverse IoT and smart‐industry environments. This hybrid design establishes a new benchmark for energy‐efficient, trustworthy, and high‐precision FL‐based anomaly detection in next‐generation WSNs.