Edge-Optimized Federated Learning for IoT Devices

Thesis Abstract

Edge-Optimized Federated Learning for IoT Devices abstracts a practical, scalable approach to harness distributed data at the network edge while addressing resource-constrained IoT environments. The study investigates how to minimize communication overhead, computation load, and energy consumption in federated learning (FL) when data resides on heterogeneous IoT devices, without sacrificing model accuracy or convergence reliability. The problem is framed around three core challenges (i) limited onboard compute, memory, and battery life on edge devices; (ii) non-IID and heterogeneous data distributions across devices; and (iii) stringent privacy and security constraints inherent to IoT ecosystems. The aim is to design, implement, and evaluate an edge-optimized FL framework that selectively aggregates updates, adapts communication frequency, and leverages lightweight models suitable for constrained devices. Specific objectives include (1) developing a tiered communication protocol that reduces uplink traffic by 60-80% through adaptive client sampling and model update compression; (2) introducing an on-device personalisation mechanism to mitigate non-IID effects via meta-learning-informed local fine-tuning; (3) evaluating a resource-aware aggregation strategy at the edge server that dynamically weights client contributions based on computational budget, data quality, and network latency; (4) implementing robust privacy-preserving techniques, including differential privacy and secure aggregation, without prohibitive performance penalties; and (5) empirically validating the framework across three IoT domains—smart homes, industrial sensors, and wearable devices—under varying network conditions. The methodology adopts a mixed-methods, design-implementation-evaluation paradigm situated in real-world IoT testbeds. The population comprises 1,200 IoT devices distributed across three deployment scenarios 400 devices in a smart-home environment, 500 industrial sensors in a manufacturing setting, and 300 wearable devices in a health-monitoring network. A stratified random sampling approach yields 200 devices for pilot experiments and 200 devices for large-scale validation, ensuring representation of heterogeneous hardware, operating systems, and data modalities. Data collection employs device-specific local datasets and synthetic augmentations to simulate non-IID distributions, with periodic ground-truth labels provided by centralized simulators to enable objective evaluation. Instruments include on-device logging for computing resource utilization, network trace collectors for communication metrics, and a centralized evaluation suite that computes accuracy, convergence speed, energy consumption, and privacy leakage indicators. The research design integrates three analytical layers (i) quantitative performance analysis using regression models and ANOVA to assess the impact of adaptive sampling, compression, and personalisation on accuracy and convergence under varying non-IID conditions; (ii) time-series analysis of convergence dynamics and communication overhead; and (iii) qualitative assessment through expert reviews of privacy and security measures and threat modelling. Analytical frameworks combine established theories to guide interpretation the No Free Lunch principle informs the generalisability constraints across heterogeneous IoT domains, while Federated Learning theory underpins aggregation and convergence guarantees. The study employs a hierarchical mixed-effects model to account for device-level and domain-level variance, and a Bayesian optimization loop to tune hyperparameters for compression rate, sampling probability, and privacy budgets. Validity and reliability are ensured through cross-validation across domains, replication of experiments with different random seeds, and sensitivity analyses on data heterogeneity and network conditions. Expected findings indicate that the proposed edge-optimized FL framework achieves comparable accuracy to centralized training with significantly reduced communication (up to 75%), lower energy expenditure (20-35% per device), and faster convergence (15-25% fewer rounds) without compromising adherence to differential privacy constraints. The meta-learning-informed personalization is anticipated to improve local model performance by 6-12% in highly non-IID settings. The study also anticipates identifying trade-offs between compression levels and model accuracy, providing actionable guidelines for deploying FL in resource-constrained IoT environments. Contribution to knowledge includes (i) a practical, scalable architecture for edge-optimized FL tailored to IoT constraints; (ii) a novel adaptive aggregation and personalization strategy guided by empirical performance data; and (iii) a comprehensive evaluation framework spanning smart-home, industrial, and wearable IoT contexts. The main conclusion posits that judicious combination of selective client participation, on-device personalization, and edge-level aggregation can deliver robust FL performance for IoT while preserving privacy and operational efficiency. Recommendations include guidelines for selecting compression schemes, privacy budgets, and sampling strategies tailored to device heterogeneity and domain requirements, plus directions for extending the framework to federated reinforcement learning for real-time decision-making at the edge.