Air-Ground Cooperative Asynchronous Online Federated Learning With Misaligned Over-The-Air Computation

In practice, air-ground cooperative federated learning (FL) assisted by unmanned aerial vehicles (UAVs) faces challenges due to lagging UAVs and the necessity of learning features from real-time data. Moreover, frequent communication and computation become performance bottlenecks for air-ground cooperative FL, especially when a large number of UAVs share the same wireless medium. Therefore, an Air-Ground Cooperative Asynchronous Online Federated Learning framework with Over-the-Air Computation (OAC-AOFL) is proposed. The OAC scheme enables UAVs to perform computations simultaneously with data transmission over a multiple-access channel. Therefore, the communication and computation delays are reduced. Additionally, a matched filter is employed to develop an aligned-sample estimator, minimizing the impact of misaligned OAC. Simulation results reveal that the proposed OAC-AOFL accelerates global model convergence speed compared to AOFL with conventional orthogonal frequency-division multiple access. Moreover, excessive noise power or degree of time misalignment can significantly impact the performance of OAC-AOFL.