OptFiD-LoRa: Optimized Federated Learning with Differential Privacy for Anomaly Detection in LoRa-IoTs: A Case Study on EDs

The increasing deployment of LoRa Internet of Things (IoT) systems in sensitive domains such as healthcare, finance, and agriculture creates a pressing need for anomaly detection methods that are simultaneously privacy-preserving, energy-aware, and deployable on resource-constrained end devices (EDs). Existing cloud-based approaches require the transmission of raw data to a central server, increasing communication overhead, latency, battery drain, and privacy exposure. Although federated learning (FL) reduces raw-data sharing, most existing FL-based anomaly detection methods still overlook the energy cost of communication rounds. Moreover, they rely on labelled attack data that is rarely available in real LoRa deployments. To address these limitations, we propose OptFiD-LoRa, an Optimised Federated Learning with Differential Privacy for Anomaly Detection framework in LoRa IoT networks. OptFiD-LoRa leverages an unsupervised neural network to learn normal transmission patterns from compact LoRa features (SF, RSSI, SNR, and payload size) directly on local EDs in FL settings. This eliminates the need for raw data transfer and labelled malicious samples. We integrate adaptive transmission control and model quantization to reduce energy consumption in communication rounds. Each ED apply a differential privacy to protect model updates during FL rounds enhancing privacy protection guarantee. Simulated experiments on real-world LoRa datasets demonstrate that OptFiD-LoRa achieves strong anomaly detection performance across heterogeneous data. The proposed optimization strategies reduce FL energy consumption by 47.54%-52.19% without materially degrading performance, maintaining strong anomaly detection (F1-scores: 0.9015-0.9985) and ROC-AUC (0.8313-0.9939) across the LoRa datasets. This demonstrates that secure and energy-aware anomaly detection is practical for real-world LoRa IoT deployments.