Towards Secure, Adaptive, and Efficient Heterogeneous Federated Learning for Resource-Constrained 6G Edge Intelligence

Author

• Sai Mani Teja Puppala, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2026

Degree Name

Doctor of Philosophy

Department

Computer Science

First Advisor

Sinha, koushik

Second Advisor

Talukder, Sajedul

Abstract

Federated learning enables distributed model training with data privacy preservation, but real-world edge deployments remain challenging due to heterogeneous data, limited device resources, unreliable networks, high communication cost, and the need for personalization. This dissertation develops secure, adaptive, and communication-efficient federated learning frameworks for resource-constrained edge environments, and contributes foundational algorithmic infrastructure for the satellite networks that increasingly underpin such systems.The dissertation makes six contributions. First, a self-regulated clustered federated learning system forms clusters using data similarity, device capability, and geographic proximity, with driver nodes performing local aggregation and dynamic health-aware recovery. Second, this architecture is extended to smart agriculture, where tractors and sensor-equipped devices operate under dynamic rural network conditions with secure model-update transmission and checkpoint-based filtering. Third, a satellite-assisted hierarchical framework for large-scale precision agriculture combines ground-based learning with LEO and GEO satellite aggregation, improving connectivity and supporting fairness-aware learning across regions with unequal network access. Fourth, VGM², a personalized federated learning framework, communicates compact geometry-based statistics rather than model weights, substantially reducing per-round communication overhead while enabling client-specific adaptation. Fifth, FN-HFL, a FractalNet-native heterogeneous federated learning architecture for satellite mega-constellations, co-designs orbital hierarchy, model depth, and agentic scheduling as a unified distributed system. Depth-heterogeneous roles are assigned across LEO, MEO, and GEO/HEO nodes, with depth-stratified aggregation defending against poisoning and radiation-induced faults. Evaluated on wildfire detection, FN-HFL achieves an AUROC of 0.891, a 53.5% reduction in inter-satellite link communication cost, and a 6.1% missed-window rate versus 29.3% for standard federated averaging. Sixth, an efficient all-pairs shortest-time path algorithm for periodically varying temporal graphs representing LEO constellations extends Hu's classical matrix multiplication scheme through a forward–backward sweep, achieving a 3.0× runtime speedup over temporal Dijkstra and a mean 6.6 snapshot-slot departure-time advantage over static routing. Overall, this dissertation advances federated learning for secure, efficient, and personalized model training across dynamic edge environments, with applications spanning smart agriculture, wildfire detection, autonomous machinery, and orbital computing.