Multi-Task Federated Learning for Unmanned Aerial Vehicle Swarms Based on Encoder-Decoder Architecture

Traditional federated learning has limitations in unmanned aerial vehicle (UAV) swarm applications, which require all participants to perform the same tasks and have the same model structure. Therefore, a multitask federated learning (M-Fed) method suitable for UAV swarms was explored, and an innovative encoder-decoder architecture was designed to enhance knowledge sharing among UAVs performing different tasks. Firstly, a direct knowledge-sharing mechanism was established for UAVs performing the same tasks, enabling effective knowledge fusion of the same tasks through direct aggregation. Secondly, for UAVs performing different tasks, the encoder parts were extracted from the encoder-decoder architectures of all UAVs to construct a global encoder. Finally, during the training process, the information from both the local encoder and the global encoder was integrated into the loss function. Iterative updates were then performed to gradually align the local decoder with the global decoder, achieving efficient cross-task knowledge sharing. Experimental results demonstrate that compared to traditional methods, the proposed method improves the performance of UAV swarms by 1.79%, 0.37%, and 2.78% on three single tasks, respectively. Although there is a slight decrease of 0.38% in performance on one task, the overall performance is still significantly increased by 2.38%. © 2024 Science Press. All rights reserved.