Joint Design of Sensing, Communication, and Computation for Multi-AAV-Enabled Over-the-Air Federated Learning

AI integrated Communication is envisaged as one of the novel usage scenarios in IMT-2030, which calls for embedding AI into network edge to offer ubiquitous intelligent services. Meanwhile, federated learning (FL) is an appealing technology for edge intelligence due to its advantages on distributed training and privacy preserving. However, FL suffers from several limitations when being executed in wireless networks, such as high communication overhead, low deployment flexibility, and the lack of analytical performance characterization. To overcome such challenges, we propose a multi-unmanned aerial vehicle (UAV)-enabled over-the-air federated learning system, where devices sense data samples for local training, and exchange model-updates with UAVs employing analog transmission scheme. Under this scheme, we first conduct convergence analysis to characterize the influence of sensing sample size, downlink broadcasting distortions, and uplink aggregation errors on the FL performance with a closed-form expression. Subsequently, a joint sensing, communication, and computation (JSCC) design problem is established to boost the FL performance among different UAVs, under the constraints of training completion time and network resource budget. We then tackle the origin intractable problem by alternatively optimizing downlink power control, uplink transceiver design, UAV deployment, and sensing batch size. Experimental evaluations demonstrate the significant performance improvement of our JSCC design compared to conventional approaches, and the effectiveness of adaptively allocating network resources across different FL rounds in suppressing the model updating errors.