Energy-Efficient Decentralized Federated Learning for UAV Swarm With Spiking Neural Networks and Leader Election Mechanism

Abstract

Federated Learning (FL) has been considered a critical technique for assisting Unmanned Aerial Vehicle (UAV) swarm to efficiently perform tasks in dynamic environments. However, deploying FL in UAV swarm is constrained by the limited energy of the UAVs and the complex communication environments within UAV swarm networks. This letter introduces a leader election-assisted Spiking Neural Networks (SNNs)-driven decentralized FL framework for UAV swarm. This framework enables UAV swarm to train a high-performance FL model while minimizing energy and time consumption, thereby enhancing real-time decision ability of UAV swarm. In particular, the SNN-driven FL allows UAV swarm to train a shared model with less energy consumption through its discrete spike event. To this end, we conduct a systematic analysis of the training challenges associated with SNN-driven FL, and we then propose an approximate derivative algorithm to address these challenges. Furthermore, we develop an intelligent leader selection scheme based on Bayes theorem designed to reduce time consumption of model parameter transmission and accelerate the model aggregation. Simulation results show that the proposed scheme outperforms baseline schemes in terms of model performance, energy and time consumption.