Federated Learning Assisted Multi-UAV Networks

Abstract

Unmanned aerial vehicles (UAVs) have been recognized as a promising technology to be used in a wide range of civilian, public and military applications. However, given their limited payload and flight time, multiple UAVs may have to be harnessed for accomplishing complex high-level tasks, where a control center can be employed for coordinating their actions. In this article, we consider image classification tasks in UAV-aided exploration scenarios, where the coordination of multiple UAVs is implemented by a ground fusion center (GFC) positioned in a strategic, but inaccessible location, such as a mountain top, where recharging the battery is uneconomical or may even be infeasible. On-board cameras are carried by each UAV and then, federated learning (FL) is invoked for reducing the communication cost between the UAVs and the GFC, and the computational complexity imposed on the GFC. In our proposed FL-aided classification approach, initially local training is performed by each UAV based on the locally collected images to create a local model. Then, each UAV sends its locally acquired model to the GFC via a fading wireless channel, where a global model is generated, which is then fed back to each UAV for the next round of their local training. In order to further minimize the computational complexity imposed on the GFC by the UAVs, weighted zero-forcing (WZF) transmit precoding (TPC) is used at each UAV based on realistic imperfect channel state information (CSI). The system performance attained is evaluated by simulations, showing that the proposed system is capable of attaining a high classification accuracy at relatively low communication cost.