SAFEL-IoT: Secure Adaptive Federated Learning with Explainability for Anomaly Detection in 6G-Enabled Smart Industry 5.0

The rise of 6G-enabled smart industries necessitates secure, adaptive, and interpretable anomaly detection frameworks capable of operating under dynamic, adversarial, and resource-constrained environments. This study presents SAFEL-IoT, a novel Secure Adaptive Federated Learning framework with integrated explainability, specifically designed for anomaly detection in Industrial Internet-of-Things (IIoT) systems under Industry 5.0 paradigms. SAFEL-IoT introduces a dynamic aggregation mechanism based on temporal model divergence, a hybrid encryption scheme combining partial homomorphic encryption with differential privacy, and an interpretable anomaly scoring pipeline leveraging SHapley Additive exPlanations (SHAP) values and temporal attention mechanisms. Extensive experimentation on the SKAB industrial dataset demonstrates that SAFEL-IoT achieves a superior F1 score of 0.93, reduces training time to 63.7 s, and maintains explanation fidelity with only a 0.15 explanation error. Communication efficiency is improved by 70.3% through 6G network slicing, while detection latency remains below 12 ms across 100 distributed edge clients. Further analysis shows a 41.7% improvement in drift robustness and a 68.9% reduction in false positives compared to traditional federated learning baselines. Theoretical convergence guarantees, scalability under large node deployments, and resilience against adversarial attacks validate SAFEL-IoT as a comprehensive and practical solution for secure, explainable, and scalable anomaly detection in next-generation industrial ecosystems.