Completion Time Minimization for UAV-Assisted Semi-Decentralized Hybrid Federated Learning

The Internet-of-Things (IoT) enables the connection of myriad wireless devices, generating a massive influx of data that can overwhelm central servers. Federated learning (FL) mitigates these issues by distributing computation tasks across edge devices, thus preserving data privacy by eliminating the need for raw data transmission. However, when deployed in large-scale IoT networks, FL encounters several challenges such as device energy constraints, heterogeneous network conditions, and the straggler effect, which inevitably impedes model convergence. In response, this paper proposes a semi-decentralized hybrid FL (SDHFL) scheme leveraging an unmanned aerial vehicle (UAV) as a mobile data center to collect data from distributed IoT clusters. To further expedite FL convergence, we formulate a non-convex mixed-integer nonlinear programming (MINLP) problem to minimize overall completion time while ensuring quality of service (QoS) requirements, considering both the UAV's energy constraints and network stability. We demonstrate that efficient learning is achieved by optimizing both power allocation and device computational capability. Furthermore, we prove the convergence of the proposed SDHFL scheme and derive the minimum number of global iterations required. Exploiting these insights, we introduce a low-complexity suboptimal algorithm for dynamic cluster selection and resource allocation optimization, leveraging Lyapunov optimization theory to obtain optimal solutions efficiently, demonstrating its scalability for large-scale networks. Our simulation results validate the effectiveness of the proposed SDHFL framework, revealing a non-trivial tradeoff where the overall completion time initially decreases and then increases as the number of devices or clusters grows, indicating the necessity of optimizing both cluster counts and intra-cluster device allocations. Furthermore, compared to several baseline schemes, our proposed optimal algorithms significantly reduce overall completion time and enhance model convergence, demonstrating the potential of SDHFL for enabling efficient and scalable FL in IoT networks.