SecuFL-IoT: an adaptive privacy-preserving federated learning framework for anomaly detection in smart industrial networks

Abstract The increasing adoption of Industrial Internet of Things (IIoT) devices introduces significant cybersecurity and privacy challenges, particularly anomaly detection and secure data sharing. This study presents SecuFL-IoT, a secure and communication-efficient federated learning framework designed for IIoT environments. SecuFL-IoT integrates adaptive anomaly detection, lattice-based homomorphic encryption, differential privacy, and reinforcement learning-based threshold adjustment to enhance security, privacy, and efficiency. The proposed model is evaluated against state-of-the-art federated learning approaches, including FedAvg, FedProx, and SCAFFOLD, using the X-IIoTID dataset. Experimental results demonstrate that SecuFL-IoT achieves an F1-score of 88.5% and a false positive rate of 2.7%, outperforming baseline models in anomaly detection accuracy. The framework reduces communication overhead by 53%, converges 23% faster than FedOPT, and lowers energy consumption by 35%, making it highly suitable for resource-constrained IIoT devices. Additionally, SecuFL-IoT ensures strong privacy guarantees ( $$\epsilon=0.9$$ ) and improves adversarial robustness, reducing data poisoning success rates below 9%. However, the framework introduces encryption latency and assumes a static network topology, which may affect real-time adaptability in highly dynamic environments. In conclusion, SecuFL-IoT provides a scalable, privacy-preserving, and industry-compliant federated learning solution that aligns with ISA/IEC 62,443 cybersecurity standards, ensuring secure anomaly detection in smart factories, power grids, and other critical IIoT infrastructures.