Semi-Supervised Federated Learning With UAV-Assisted Data Acquisition in Wireless Networks
Abstract
Federated learning (FL) can effectively enable multiple distributed devices to collaboratively train a global model without sharing raw data, offering strong data privacy protection and scalability for the Internet of Things (IoT). However, FL typically requires a large amount of labeled data to achieve global model convergence, which limits its applicability in real-world scenarios. To address this issue, we propose a semi-supervised federated learning (SSFL) framework for uncrewed aerial vehicle (UAV)-enabled surveillance systems. Each UAV collects unlabeled surveillance data and trains a local model by adopting both pre-stored labeled data and newly collected unlabeled data. To minimize the total completion time of the SSFL process, we first analyze the correlation between the convergence of the global model and the volume of unlabeled data. We further reveal a tradeoff between the number of communication rounds and the per-round training latency. Moreover, we derive that there exists an optimal oblique photography (OP) angle for UAV-based data collection, which simultaneously improves image resolution and reduces the transmission latency. These findings highlight the unique interplay between sensing geometry and communication efficiency, which has been largely overlooked in previous works. Finally, we jointly optimize the UAV position, OP angle, unlabeled data volume, resource allocation, and number of communication rounds to minimize the SSFL completion time in a multi-UAV network. Simulation results show that compared with existing benchmarks, the proposed scheme significantly reduces training time and accelerates convergence speed. Moreover, our results reveal two key tradeoffs: (1) the amount of unlabeled data must balance local computation delay against global convergence gain, and (2) the OP angle governs image fidelity and transmission efficiency, reflecting the intrinsic linkage between sensing and communication design.