| Abstrakt: |
In Federated Learning (FL), varying local dataset distributions among clients can result in unsatisfactory performance for some, leading to an unfair model. While some prior works attempted to resolve this issue, their approaches fall short of training a fair model in real-world scenarios where imperfections in wireless channels cause the server to receive noisy versions of local updates. To tackle this issue, in this letter, we treat FL as a multi-objective minimization problem, and develop a fair FL algorithm that explicitly accounts for the inherent imperfections of wireless channels. Particularly, we modify the classical multiple gradient descent algorithm to assist the server in identifying a common descent direction for all local objectives based on the received noisy gradients. Additionally, we evaluate the performance of the proposed method via some experiments. |
