Adaptive Privacy-Preserving Split-Hierarchical Federated Learning for Resource-Constrained IoT Networks

The proliferation of Internet-of-Things (IoT) devices necessitates efficient machine learning paradigms that address bandwidth constraints, privacy requirements, and computational heterogeneity. While hierarchical federated learning offers com- munication efficiency and split learning reduces computational burden on resource-constrained devices, existing approaches lack adaptive mechanisms for dynamic environments and formal privacy guarantees. We propose AP-SHFL (Adaptive Privacy- Preserving Split-Hierarchical Federated Learning), a novel three- tier architecture that jointly optimizes split point selection, hierarchical aggregation, and differential privacy mechanisms. Our approach employs Q-learning for per-client dynamic split point adaptation based on real-time loss and communication feed- back, while implementing staleness-adaptive differential privacy that calibrates noise injection according to model freshness in asynchronous settings. Experimental results on MNIST demon- strate 99.19% test accuracy, 40-60% communication reduction compared to FedAvg baseline, rapid convergence (>98.5% in <10 rounds), and adaptive split point evolution from an average of 1.10 to 2.75 across clients, showcasing effective per-client optimization. Ablation studies confirm the contribution of each component. Our framework achieves state-of-the-art results, outperforming HSFL (98.1% accuracy) while maintaining formal differential privacy guarantees.