Adaptive Hierarchical Federated Learning for IoT Anomaly Detection

Abstract

The rapid growth of Internet of Things (IoT) networks introduces unprecedented security challenges, particularly in detecting and responding to evolving cyber threats in a scalable, efficient, and privacy-preserving manner. This paper presents AHFL-DAWA, a novel Adaptive Hierarchical Federated Learning framework with Dynamic Anomaly-Weighted Aggregation, designed to enhance threat detection in large-scale IoT environments. AHFL-DAWA integrates dynamic anomaly scoring, hierarchical model aggregation, and formal differential privacy guarantees to achieve high detection accuracy while preserving data confidentiality and communication efficiency. Experimental results across diverse datasets and attack scenarios show that AHFL-DAWA achieves 96.8% accuracy—surpassing centralized and traditional FL baselines—while reducing communication overhead by 78.4% and energy consumption by 34.2%. The framework also demonstrates strong robustness under adversarial conditions, maintaining 94.3% accuracy with up to 29.3% device compromise. Comprehensive evaluations confirm its scalability to millions of devices, resilience to Byzantine faults, and minimal degradation under non-IID data distributions. These findings position AHFL-DAWA as a practical and secure solution for federated intrusion detection in next-generation IoT systems.