Date of Defense
8-4-2026 11:00 AM
Location
Microsoft Teams
Document Type
Dissertation Defense
Degree Name
Doctor of Philosophy (PhD)
College
College of Information Technology
Department
Information Systems and Security
First Advisor
Zouheir Trabelsi
Keywords
Federated learning, Hierarchical fog computing, Dynamic aggregation placement, Decentralized coordination, Post-quantum cryptography
Abstract
Federated learning (FL) enables collaborative model training without centralizing raw data, but deploying FL at scale in real edge environments remains challenging because iterative training and aggregation must operate over heterogeneous, resource-constrained, and often mobile devices with time-varying connectivity. Conventional hierarchical federated learning (HFL) partially mitigates communication cost by introducing fog/edge aggregation, yet many designs retain cloud-based global aggregation and cloud-centric coordination. This places wide-area network latency on the critical path of every training round, creates a single point of failure, and limits responsiveness as model sizes and federation scale grow. Moreover, moving coordination and aggregation closer to the edge enlarges the attack surface, increasing the importance of strong integrity, authenticity, and long-term confidentiality guarantees, including resilience against future quantum-capable adversaries. This thesis develops an end-to-end framework that progressively transforms cloud-centric HFL into cloud-minimized and fully decentralized edge learning, and then secures its coordination plane with post-quantum cryptography while keeping overhead practical through cryptographic agility. First, the thesis establishes a systems foundation for scalable edge learning via a mobility- and resource-aware hierarchical fog framework. The framework dynamically promotes nearby devices with available compute and storage into fog roles under explicit constraints, and organizes coordination through multi-level fog tiers and brokers. This improves utilization, reduces bottlenecks and stragglers, and stabilizes iterative learning workflows under mobility and intermittent links. Building on this substrate, the thesis proposes a fog-centric HFL architecture with dynamic global aggregation placement. Global aggregation is moved from the cloud to a fog node selected adaptively using multi-criteria optimization over communication cost, workload, availability, reliability, and aggregated data-quality and contribution indicators. The cloud is retained only as a lightweight coordinator outside the aggregation critical path. This design reduces end-to-end round latency and improves training efficiency by eliminating the mandatory cloud round trip while avoiding persistent overload at any single fog node. To remove centralized dependence entirely, the thesis then introduces a fully decentralized hierarchical FL architecture in which fog/edge nodes form an authenticated peer-to-peer overlay and elect a per-round global aggregator using gossip dissemination and consensus-based selection, with fault-aware re-election to sustain progress under failures and churn. This decentralization improves autonomy and resilience in environments with intermittent or unavailable cloud connectivity, while preserving quality-aware aggregation and communication-efficient synchronization. The thesis further develops post-quantum secure coordination mechanisms to protect update exchange and control-plane operations. It integrates post-quantum signatures for end-to-end authenticity, accountability, and non-repudiation, and post-quantum key encapsulation mechanisms (KEMs) to establish secure channels whose steady-state traffic is protected by symmetric authenticated encryption. Optional ledger support provides tamper-evident logging and auditability with scalable partitioning and anchoring. Finally, to reduce the variable overhead of PQ key establishment, the thesis introduces an adaptive, learning-based PQ-KEM selector trained on benchmark measurements and evaluated under realistic fog/vehicular conditions with queueing effects. The selector recommends an efficient KEM satisfying a required security level and authenticates the recommendation to prevent manipulation during first contact. Overall, the thesis demonstrates that aggregation placement, coordination strategy, and cryptographic configuration can be jointly designed to achieve low-latency, robust, and post-quantum-ready distributed learning at the edge, and provides design guidelines and limitations to inform practical deployments.
Included in
Fully Decentralized Hierarchical Federated Learning at the Edge with Post-Quantum Secure Communication
Microsoft Teams
Federated learning (FL) enables collaborative model training without centralizing raw data, but deploying FL at scale in real edge environments remains challenging because iterative training and aggregation must operate over heterogeneous, resource-constrained, and often mobile devices with time-varying connectivity. Conventional hierarchical federated learning (HFL) partially mitigates communication cost by introducing fog/edge aggregation, yet many designs retain cloud-based global aggregation and cloud-centric coordination. This places wide-area network latency on the critical path of every training round, creates a single point of failure, and limits responsiveness as model sizes and federation scale grow. Moreover, moving coordination and aggregation closer to the edge enlarges the attack surface, increasing the importance of strong integrity, authenticity, and long-term confidentiality guarantees, including resilience against future quantum-capable adversaries. This thesis develops an end-to-end framework that progressively transforms cloud-centric HFL into cloud-minimized and fully decentralized edge learning, and then secures its coordination plane with post-quantum cryptography while keeping overhead practical through cryptographic agility. First, the thesis establishes a systems foundation for scalable edge learning via a mobility- and resource-aware hierarchical fog framework. The framework dynamically promotes nearby devices with available compute and storage into fog roles under explicit constraints, and organizes coordination through multi-level fog tiers and brokers. This improves utilization, reduces bottlenecks and stragglers, and stabilizes iterative learning workflows under mobility and intermittent links. Building on this substrate, the thesis proposes a fog-centric HFL architecture with dynamic global aggregation placement. Global aggregation is moved from the cloud to a fog node selected adaptively using multi-criteria optimization over communication cost, workload, availability, reliability, and aggregated data-quality and contribution indicators. The cloud is retained only as a lightweight coordinator outside the aggregation critical path. This design reduces end-to-end round latency and improves training efficiency by eliminating the mandatory cloud round trip while avoiding persistent overload at any single fog node. To remove centralized dependence entirely, the thesis then introduces a fully decentralized hierarchical FL architecture in which fog/edge nodes form an authenticated peer-to-peer overlay and elect a per-round global aggregator using gossip dissemination and consensus-based selection, with fault-aware re-election to sustain progress under failures and churn. This decentralization improves autonomy and resilience in environments with intermittent or unavailable cloud connectivity, while preserving quality-aware aggregation and communication-efficient synchronization. The thesis further develops post-quantum secure coordination mechanisms to protect update exchange and control-plane operations. It integrates post-quantum signatures for end-to-end authenticity, accountability, and non-repudiation, and post-quantum key encapsulation mechanisms (KEMs) to establish secure channels whose steady-state traffic is protected by symmetric authenticated encryption. Optional ledger support provides tamper-evident logging and auditability with scalable partitioning and anchoring. Finally, to reduce the variable overhead of PQ key establishment, the thesis introduces an adaptive, learning-based PQ-KEM selector trained on benchmark measurements and evaluated under realistic fog/vehicular conditions with queueing effects. The selector recommends an efficient KEM satisfying a required security level and authenticates the recommendation to prevent manipulation during first contact. Overall, the thesis demonstrates that aggregation placement, coordination strategy, and cryptographic configuration can be jointly designed to achieve low-latency, robust, and post-quantum-ready distributed learning at the edge, and provides design guidelines and limitations to inform practical deployments.