Air-Ground Integrated Online Federated Learning Under Unreliable Communication
Yuqian Jing, Yuben Qu, Tao Wu, Chao Dong, Song Guo, Qihui Wu · IEEE Transactions on Cognitive Communications and Networking · 2025
In recent years, autonomous aerial vehicles (AAVs) equipped with intelligent computing modules for machine learning (ML) have garnered attention. Federated learning (FL), as an emerging distributed learning paradigm, aligns seamlessly with this scenario. AAVs train local models while a terrestrial base station (BS) aggregates the global model, forming an air-ground integrated federated learning (AGIFL) system. However, under air-ground integrated networks with unreliable wireless links, the online retraining of ML models with new samples is promising but challenging in AGIFL. To this end, we study how to enhance the performance of air-ground integrated online FL (AGIOFL) considering packet transmission error. Specifically, we formulate a joint optimization problem for sample selection and client scheduling, aiming to minimize the loss function that maximizes the performance of AGIOFL. To solve the aforementioned problem without explicit expression, we first derive a closed-form expression for the expected convergence rate of the FL algorithm to quantify the impact of AGIOFL and unreliable wireless transmission. Based on this relationship, we reformulate the optimization problem into a mixed-integer nonlinear programming problem. Subsequently, we develop an alternating iterative optimization algorithm (OFL-PE) combining branch-and-bound with the Dinkelbach algorithm to solve this problem. Simulations indicate that OFL-PE converges within 10 iterations only, with a maximum error of less than 7% compared to the exhaustive brute force approach. In addition, we construct a proof-of-concept system based on widely used AAV embedded computers, and experimental results show that OFL-PE demonstrates superior accuracy, convergence, and robustness compared to three existing FL algorithms.