Self-Adaptive Clustering Based Federated Learning Scheme in UAV Swarm Network

Federated learning (FL) as a kind of distributed machine learning paradigm enables unmanned aerial vehicles (UAVs) to conduct some novel applications, such as remote sensing and target recognition. However, high mobility of UAVs results in link quality fluctuating rapidly which influences the performance of FL models regarding convergence rate and model accuracy. In this paper, a self-adaptive clustering FL scheme in UAV swarm network is proposed to extend the lifetime of the network and improve the model performance of FL. UAVs can be dynamic clustered without predefining the number of clusters during each iteration and the cluster head selection criterion is proposed to make aggregation for each cluster. Meanwhile, to protect data privacy, the cluster gateway and an UAV leader are selected to be responsible for parameters transmission and making global aggregation respectively. Simulation results show that the proposed scheme significantly increases the network lifetime and gains a better training performance compared to the traditional FL applied in UAVs.