A Scalable Framework for Real-Time Edge-Aware Federated Learning Systems

Thesis Abstract

In an era of ubiquitous intelligent devices and latency-sensitive applications, real-time federated learning (FL) faces challenges of scalability, communication efficiency, data heterogeneity, and dynamic edge environments where devices frequently join and depart. This study addresses the problem of designing a scalable, edge-aware FL framework capable of real-time model updates without compromising privacy or performance. The aim is to develop a cohesive framework that orchestrates cross-edge collaboration, optimizes communication-computation trade-offs, and adapts to non-IID data distributions and fluctuating network conditions. Specific objectives include (1) formulating an edge-aware aggregation mechanism that balances global accuracy with per-edge latency constraints; (2) devising a dynamic participation and resource-allocation scheme based on network topology, device reliability, and data relevance; (3) designing a model-quantization, compression, and asynchronous update protocol to minimize bandwidth usage; (4) integrating differential privacy and secure aggregation to uphold privacy guarantees under real-time constraints; (5) evaluating the framework across synthetic and real-world datasets with varying degrees of heterogeneity and mobility. The methodology employs a mixed-methods research design combining quantitative experimentation with a qualitative interpretation of system dynamics. The population comprises edge devices (simulated and real) within a metropolitan-scale testbed aggregating 1,000 to 5,000 participating nodes across heterogenous hardware. A stratified random sample of 200 edge devices is selected for controlled experiments, with additional simulations extending to 1,000 and 10,000 virtual nodes to assess scalability limits. Data collection instruments include (i) device telemetry capturing latency, bandwidth, CPU/GPU utilization, and energy consumption; (ii) labeled and unlabeled sensor datasets representing traffic, environmental sensing, and multimedia data; and (iii) synthetic fault injection scenarios to model device churn and network perturbations. Analytical techniques encompass hierarchical Bayesian modeling to estimate server-client latency distributions and convergence properties; teleporting-average and FedAvg-inspired aggregation with adaptive weighting to handle non-IID data; stochastic gradient descent with momentum for optimization; dynamic resource-allocation optimization using mixed-integer linear programming; and privacy-preserving assessments through differential privacy accounting and secure multiparty computation performance benchmarks. The study adopts the theoretical lens of federated optimization theory and edge computing governance, drawing on the theories of multi-agent systems and information bottleneck to justify the design of adaptive aggregation, model compression, and selective participation strategies. Key expected findings include (i) a scalable, edge-aware aggregation scheme that reduces total training time by 40–60% under constrained bandwidth while maintaining model accuracy within 1–2 percentage points of centralized benchmarks on benchmark datasets such as CIFAR-10 and FEMNIST; (ii) evidence that dynamic participation improves convergence speed and resilience to churn, with performance gains of 15–25% relative to static participation baselines; (iii) a robust asynchronous update protocol achieving near-synchronous convergence in non-IID settings with device heterogeneity; (iv) quantifiable privacy-utility trade-offs demonstrating that differential privacy parameters can be tuned to satisfy regulatory requirements without significantly degrading accuracy; and (v) comprehensive scalability analyses showing framework viability up to 10,000 virtual nodes and real-time inference latency under 100 ms for typical edge tasks. This study contributes to knowledge by delivering a principled, scalable framework for real-time edge-aware FL, integrating adaptive participation, dynamic resource allocation, and privacy-preserving communication into a unified architecture. It advances federated optimization theory by operationalizing edge-aware convergence criteria and by bridging theoretical guarantees with practical system design for real-time contexts. The findings inform guidelines for deploying FL at metropolitan scales, including network provisioning, QoS-aware scheduling, and privacy parameterization. The study concludes with recommendations for standardizing edge-aware FL protocols, outlining opportunities for extending the framework to include reinforcement-learning-based controller adaptation and cross-domain federation among heterogeneous edge domains.