Boyu Fan defends his PhD thesis on Federated Learning in Heterogeneous Systems: From Internet of Things to Foundation Models

M.Eng. Boyu Fan defends his PhD thesis “Federated Learning in Heterogeneous Systems: From Internet of Things to Foundation Models” on Friday the 5th of June 2026 at 12 in the University of Helsinki Exactum building, Auditorium CK112 (Pietari Kalmin katu 5, basement). His opponent is Professor Hamed Haddadi (Imperial College London, United Kingdom) and custos Professor Pan Hui (University of Helsinki). The defence will be held in English.

The thesis of Boyu Fan is a part of research done in the Department of Computer Science and in the Systems and Media group at the University of Helsinki. His supervisor has been Professor Pan Hui (University of Helsinki).

Federated Learning in Heterogeneous Systems: From Internet of Things to Foundation Models

Federated Learning (FL) has emerged as a cornerstone of edge intelligence by enabling privacy-preserving and collaborative training of AI models on distributed nodes without exposing sensitive raw data. However, the practical deployment of FL in real-world Internet of Things (IoT) environments is severely hindered by the heterogeneity gap, i.e., the profound disparity in device hardware capabilities and the diversity of local data distributions. Traditional FL protocols, which rely on rigid synchronous aggregation and ``one-size-fits-all'' model architectures, fail to scale effectively in such highly dynamic environments. This rigidity leads to severe straggler effects, where system performance is strictly bounded by the weakest node, and often results in model drift due to statistical divergence. Consequently, existing approaches struggle to balance training efficiency with model accuracy, particularly when scaling to resource-constrained edge devices.

To systematically address these challenges, this dissertation presents a systematic progression of heterogeneity-aware methodologies that evolves from communication-centric optimizations to model-centric adaptations. First, we tackle communication bottlenecks in IoT networks by proposing SAFI, a semi-asynchronous protocol that isolates stragglers through latency-based clustering and a hybrid synchronization mechanism. Second, moving to the model architecture level, we dismantle the structural rigidity of standard FL with FedTSA and FedMixer. These frameworks empower resource-limited clients to train customized sub-models tailored to their hardware budgets and temporal patterns, utilizing novel generative and mutual Knowledge Distillation (KD) protocols to bridge the dimensionality gap. Finally, tackling the unique barriers to deploying Foundation Models (FMs) on resource-constrained devices, we present HeLoRA. Focusing on the heterogeneity issues within Parameter-Efficient Fine-Tuning (PEFT), this framework resolves the dimensionality mismatch of diverse LoRA adapters through context-aware structural alignment and consensus-driven KD.

The significance of this research lies in orchestrating a systematic evolution in FL: advancing from efficient synchronization protocols to structure-agnostic knowledge fusion. By enabling architectural diversity aligned through knowledge transfer, we demonstrate that system efficiency and model accuracy are not mutually exclusive but can be co-optimized. This work effectively democratizes access to advanced AI, allowing devices ranging from low-power sensors to high-end edge servers to collaborate seamlessly. Ultimately, the proposed heterogeneity-aware frameworks pave the way for a more inclusive, efficient, and scalable ecosystem for ubiquitous edge intelligence, capable of supporting the next generation of personalized and privacy-preserving AI applications.

Avail­ab­il­ity of the dis­ser­ta­tion

An electronic version of the doctoral dissertation will be available in the University of Helsinki open repository Helda at .

Printed copies will be available on request from Boyu Fan: .