Abstract

UAV swarms are widely used in emergency communications, area monitoring, anddisaster relief. Coordinated by control centers, they are ideal for federatedlearning (FL) frameworks. However, current UAV-assisted FL methods primarilyfocus on single tasks, overlooking the need for multi-task training. Indisaster relief scenarios, UAVs perform tasks such as crowd detection, roadfeasibility analysis, and disaster assessment, which exhibit time-varyingdemands and potential correlations. In order to meet the time-varyingrequirements of tasks and complete multiple tasks efficiently under resourceconstraints, in this paper, we propose a UAV swarm based multi-task FLframework, where ground emergency vehicles (EVs) collaborate with UAVs toaccomplish multiple tasks efficiently under constrained energy and bandwidthresources. Through theoretical analysis, we identify key factors affecting taskperformance and introduce a task attention mechanism to dynamically evaluatetask importance, thereby achieving efficient resource allocation. Additionally,we propose a task affinity (TA) metric to capture the dynamic correlation amongtasks, thereby promoting task knowledge sharing to accelerate training andimprove the generalization ability of the model in different scenarios. Tooptimize resource allocation, we formulate a two-layer optimization problem tojointly optimize UAV transmission power, computation frequency, bandwidthallocation, and UAV-EV associations. For the inner problem, we deriveclosed-form solutions for transmission power, computation frequency, andbandwidth allocation and apply a block coordinate descent method foroptimization. For the outer problem, a two-stage algorithm is designed todetermine optimal UAV-EV associations. Furthermore, theoretical analysisreveals a trade-off between UAV energy consumption and multi-task performance.