Edge-Aware Federated Learning for IoT Networks with Heterogeneous Data

Thesis Abstract

Edge-Aware Federated Learning (EA-FL) is explored to address the challenges of training robust machine learning models in Internet of Things (IoT) ecosystems characterized by heterogeneous data distributions, limited device resources, and variable network conditions. The problem addressed centers on performance degradation caused by non-IID data, device stragglers, and communication bottlenecks that impede model convergence and accuracy in standard federated learning (FL) frameworks. The aim is to develop an EA-FL framework that (i) adapts aggregation and model personalization to local data heterogeneity, (ii) optimizes communication efficiency through edge-assisted scheduling and compression, and (iii) enhances resilience to stragglers and unreliable links through asynchronous updates and fault-tolerant mechanisms. Specific objectives include (1) designing an edge-aware aggregation rule that weights client updates by data heterogeneity metrics and local model drift; (2) integrating data-characteristic-driven personalization layers to improve local accuracy without sacrificing global generalization; (3) developing a communication-efficient pipeline incorporating gradient sparsification and adaptive compression at the edge gateway; (4) evaluating robustness against heterogeneity and device dropouts using synthetic and real-world IoT datasets; (5) benchmarking against baseline FL, FedAvg, and personalized FL variants; and (6) deriving guidelines for deployment in smart city and industrial IoT scenarios. The methodology adopts a mixed-methods, quasi-experimental research design combining quantitative evaluation with qualitative performance profiling. The population comprises IoT devices and edge gateways deployed in a simulated smart building and industrial setting, totaling 1,000 heterogeneous devices distributed across three edge servers. A stratified random sample of 300 devices is selected for experiments, ensuring representation across device capabilities, data distributions, and network reliability. Data collection uses three sources (i) synthetic IoT time-series datasets with controlled non-IID properties (e.g., varying label distributions and feature correlations), (ii) publicly available IoT benchmarks (e.g., the UCI Appliance Energy dataset, IoT-23), and (iii) field traces from a pilot deployment of a campus smart building. Instruments include a modular EA-FL software stack implemented in PyTorch with edge scheduling enabled by Kubernetes, device telemetry dashboards, and logging for convergence metrics. Validity and reliability are addressed through cross-validation (k=5) for predictive tasks, pre-registration of experiments, and repeatability via containerized experiments. The data analysis employs (a) iterative optimization to calibrate the edge-aware aggregation weights using a Bayesian hierarchical model, (b) statistical tests including repeated-measures ANOVA to compare convergence speed and accuracy across FL variants, and (c) ablation studies to quantify the impact of personalization layers, communication compression, and edge scheduling. Model evaluation uses accuracy, F1-score, communication rounds to convergence, and energy consumption as key metrics. The EA-FL framework is evaluated on two primary tasks (i) anomaly detection in IoT telemetry and (ii) predictive maintenance for edge-enabled industrial sensors. Expected findings indicate that the edge-aware aggregation improves global accuracy by 6–12% over FedAvg under highly non-IID conditions, reduces total communication by 25–40% via adaptive compression, and accelerates convergence by 20–30% in comparison to baseline FL. Personalization layers are anticipated to yield a local accuracy uplift of 4–9% without disproportionately increasing model drift, while asynchronous updates and straggler-robust mechanisms are expected to maintain stable performance with up to 20% device dropout. The study contributes to knowledge by (i) formalizing an edge-aware aggregation scheme that jointly optimizes heterogeneity handling and communication efficiency, (ii) integrating data-driven personalization within FL for IoT contexts, and (iii) providing a practical blueprint for deploying EA-FL in real-world smart environments with limited compute and bandwidth. Theoretical grounding leverages the Communication-Efficiency theory in FL and the Heterogeneity Robustness framework, supplemented by concepts from Federated Personalization and Edge Computing. The main conclusion posits that edge-aware, asynchronous federated learning with selective personalization substantively enhances model performance and resource efficiency in heterogeneous IoT networks. Recommendations include adopting EA-FL in smart city deployments, prioritizing edge-level data characterization to determine aggregation weights, and extending the framework to multi-tenant IoT ecosystems with privacy-preserving mechanisms such as secure aggregation. Suggestions for future research include exploring transfer learning between heterogeneous domains, dynamic edge provisioning responsive to network conditions, and extending the theoretical model to non-stationary data streams.