Self-adaptive federated learning for scalable IoT applications on resource-constrained edge devices

Aljohani, Abdulaziz 2026. Self-adaptive federated learning for scalable IoT applications on resource-constrained edge devices. PhD Thesis, Cardiff University. Item availability restricted.

	PDF - Accepted Post-Print Version Restricted to Repository staff only until 18 September 2027 due to copyright restrictions. Download (22MB)
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (261kB)

Abstract

The rapid growth of the Internet of Things (IoT) has shifted machine learning deployment from centralised cloud infrastructure to resource-constrained edge devices. Local inference reduces latency and communication overhead and can enhance privacy, but deployed models often remain static and become vulnerable to concept drift, evolving environments, and changing device behaviour. Post-deployment adaptation is restricted by limited labelled data and tight compute, memory, and energy budgets. Federated learning offers privacy-preserving collaboration across devices, but practical IoT deployments must address both statistical heterogeneity (non-IID data) and system heterogeneity (uneven resources and connectivity). This thesis investigates self-adaptive federated learning for continuous on-device adapta tion in IoT. It formalises a self-adaptive Internet of Federated Things (IoFT) paradigm and derives a multi-dimensional taxonomy and a MAPE-K-based conceptual architecture for post-deployment learning. To mitigate the lack of labels, the thesis introduces TinyLearnED, a weakly supervised anomaly detection framework that exploits cross-device correlations to generate pseudo-labels for on-device training. Furthermore, it is coupled with a lightweight and robust federated aggregation strategy to stabilise learning under anomalous updates while updating a constrained parameter set. To improve scalable personalisation under statistical heterogeneity, the thesis presents ReactFed, which combines adaptive client clus tering with dynamic structured sparsification to deliver cluster-specific submodels with resource-aware capacity and includes a stability mechanism to reduce forgetting when clients transition between clusters. Experiments on real-world and benchmark evaluations show that TinyLearnED improves anomaly detection and that ReactFed enhances personalised performance while reducing resource overheads, demonstrating the feasibility of continuous adaptation on constrained edge devices.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software
Date of First Compliant Deposit:	18 September 2026
Last Modified:	22 Sep 2026 15:13
URI:	https://orca.cardiff.ac.uk/id/eprint/189708

Actions (repository staff only)

Edit Item