Joint Data Retransmission and Client Selection Optimization for Error-Tolerant Federated Learning in AAV Networks

The flexibility and mobility of autonomous aerial vehicles (AAV) swarms enable them to integrate with federated learning (FL), an emerging distributed machine learning framework. AAV-FL creates an edge intelligence system for AAV swarms. However, unreliable AAV networks lead to a large amount of retransmissions for the dropped data, making it difficult for FL in AAV swarms to achieve the expected accuracy within a short period. Fortunately, benefiting from the dilution effect of convolutional computations, FL can tolerate a limited amount of model parameter errors. Thus motivated, this paper proposes a novel FL framework called FedET, to jointly optimize data retransmission and client selection to achieve error-tolerant FL in AAV networks. Specifically, FedET utilizes error tolerance of FL to reduce the number of retransmissions, which navigates the trade-off between training time and model accuracy. Meanwhile, the impact of retransmission on client selection is also analyzed. We formulate the training utility maximization problem for FL via jointly optimizing data retransmission and client selection. To solve this problem, we propose an alternating optimization-based algorithm to reach the local optimal solution. We implement and evaluate FedET on widely used real-world AAV embedded devices (i.e., NVIDIA Jetson devices). Compared to existing FL algorithms, when faced with unreliable AAV networks, FedET on average reduces the total training time by ~45.8%, and transmission latency by ~67.9%, respectively.