Year: 2026
Type: dissertation
Abstract: Federated learning (FL) enables distributed clients to train a shared model by exchanging model information over communication networks while retaining raw data locally. Beyond learning algorithms, its operation depends on how data are distributed, how clients are connected, and what constraints or disruptions they encounter. This thesis investigates robustness and resilience in federated learning over networked systems: maintaining effective learning under constraints and disruptions and restoring effective operation after disruption. It addresses four practical constraints in centralized FL for battery-based energy storage systems: early-deployment scarcity of abnormal data, model exchange overhead, malicious model updates, and packet loss during model transmission. For decentralized federated learning (DFL), this thesis proposes a topology with a bimodal degree distribution to preserve network connectivity under random failures and targeted attacks and introduces teleportation links to accelerate information propagation and mitigate catastrophic forgetting under heterogeneous data. Vulnerability analysis uses topology and data distribution to examine where and why DFL networks are vulnerable. Recovery planning addresses how to restore access to distributed information after network disruption. Finally, this thesis introduces Effective Data-Oriented Community (DOC) Diversity to characterize node-level information accessibility. Experimental evaluations of these contributions show how federated learning can remain effective under practical constraints and recover from network disruptions.
Author: Wang, Xu
Institution: Queen's University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: federated learning, decentralized federated learning, network resilience, network robustness, networked systems, battery energy storage, packet loss, network topology, random failures, targeted attacks, catastrophic forgetting, data heterogeneity, information dissemination, vulnerability assessment
Open Access status: green
OpenAlex ID: W7225842752
