Edge-Optimized Federated Learning for Resource-Constrained IoT Networks

Thesis Abstract

Edge-Optimized Federated Learning (EOFL) addresses the critical challenge of deploying machine learning in resource-constrained IoT networks where data privacy, limited computational power, and intermittent connectivity hinder centralized training. The study investigates how to design, implement, and evaluate a federated learning framework that minimizes on-device computation, optimizes communication efficiency, and preserves model accuracy in heterogeneous edge environments. The aim is to develop an EOFL architecture that (i) reduces energy consumption and latency on IoT devices, (ii) sustains robust model performance across diverse data distributions, and (iii) provides secure and privacy-preserving collaborative learning suitable for real-world deployments. Specific objectives include (1) to analyze the trade-offs between local training workload, communication rounds, and global model convergence under varying device capabilities; (2) to design an adaptive federation scheme that leverages gradient sparsification, quantization, and selective client participation to reduce communication overhead while maintaining accuracy; (3) to integrate differential privacy and secure aggregation techniques to ensure data confidentiality within resource limits; (4) to develop a lightweight on-device inference engine that leverages edge acceleration while minimizing energy use; and (5) to empirically evaluate EOFL against baseline federated learning and centralized training across multiple IoT scenarios with heterogeneous data sources. The research adopts a design, implementation, and evaluation methodology anchored in an experimental, quasi-experimental design. The population comprises IoT edge devices and gateways deployed in three representative environments smart home sensor networks, industrial IoT in manufacturing floors, and urban environmental monitoring systems. A stratified random sample of 60 devices per environment is selected, totaling 180 devices, with 40% representing low-power microcontrollers, 40% mid-range single-board computers, and 20% edge gateways. Data collection uses benchmarked public datasets restructured to reflect non-IID distributions typical of IoT contexts, including energy, environmental, and activity recognition data. Instruments include a local training pipeline simulator on-device, a network simulator for edge-to-edge communication, and a centralized orchestration console for benchmarking. Validity and reliability are ensured through cross-device calibration, repeatable data partitions, and replication of experiments across three randomized seeds. Analytical methods combine quantitative and qualitative approaches. Model performance is assessed using accuracy, F1-score, energy consumption (mJ per training step), training time (seconds per global round), and communication cost (bytes per round). Regression analysis and multivariate ANOVA quantify the impacts of local computation, communication strategies, and device heterogeneity on convergence speed and accuracy under non-IID data. An ablation study isolates the contributions of gradient sparsification, quantization, and selective participation. Privacy guarantees are evaluated via membership inference risk estimation and differential privacy budget tracking. The study employs a theoretical lens grounded in Distributed Optimization Theory and the Information Bottleneck principle to justify the trade-offs between compression, privacy, and model utility. The EOFL framework is implemented in PyTorch with custom communication protocols and evaluated against two baselines standard FedAvg with uniform participation and a centralized model trained on aggregated data. Thematic analysis of engineering notes and logs supports qualitative insights into system reliability, scalability, and usability. Expected findings indicate that EOFL can achieve near-baseline accuracy (within 2–3 percentage points on average) with 35–50% reductions in communication load and 20–40% energy savings on low-power devices, while maintaining stable convergence under non-IID data. The adaptive participation mechanism is anticipated to yield faster convergence in heterogeneous networks and improve model robustness against straggler devices. The integration of differential privacy is expected to incur a modest privacy-utility trade-off, with configurable privacy budgets enabling application-specific guarantees. The study contributes to knowledge by providing a rigorously validated blueprint for deploying privacy-preserving, computation-, and communication-efficient federated learning in real-world IoT ecosystems, supported by reproducible experimental data and open-source tooling. The main conclusion posits that edge-optimized federated learning offers a viable pathway for scalable, privacy-preserving collaborative learning in resource-constrained IoT networks, achieving meaningful accuracy with substantial energy and bandwidth savings. Recommendations include adopting adaptive participation and gradient compression as standard practice in EOFL deployments, balancing privacy budgets with application requirements, and continuing to explore hardware-aware optimizations for future IoT architectures.