Mutil-UAV Collaborative Split Learning for Frequency-Hopping Sequence Prediction

In complex electromagnetic environments, such as electromagnetic warfare scenarios, unmanned aerial vehicles (UAVs) usually need to update frequency-hopping (FH) sequence prediction models using collected data at low altitudes. However, such tasks generally consume substantial computational resources, severely straining the limited memory capacity of UAV platforms. While some existing works employ split learning to address the training limitations on resource-constrained devices like UAVs, these methods primarily focus on feedforward neural networks and lack effective partitioning strategies for recurrent neural networks in FH sequence prediction tasks. To address this challenge, this paper proposes the Multi-UAV collaborative Split Learning (MUSL) algorithm. First, the auxiliary UAV determine partition points based on the computational capabilities within the UAV swarm. Then, the auxiliary UAV and client UAVs perform efficient split training based on sequences rather than hidden layers. Finally, the auxiliary UAV periodically aggregate the models. We successfully deployed the MUSL algorithm on real-world UAV onboard computing platforms. Experimental results show that MUSL reduces the average memory usage by ~7.21% compared to centralized learning, and shortens average training latency by ~51.87% compared to federated split learning.