Signal Processing Optimization for Federated Learning over Multi-User MIMO Uplink Channel

In federated learning, remote mobile devices, which are equipped with local datasets, collaborate through a parameter server (PS) in order to train a machine learning model. An advantage of the federated learning is its effectiveness of preserving the privacy of local raw data. However, it is challenging to meet the demands on latency of exchanging data on wireless multiple access channel (MAC) with limited bandwidth. Over-the-air computation (AirComp) is a potential solution to this problem, which leverages the superposition property of MAC channel. This work addresses the signal processing optimization of both digital federated learning and AirComp schemes under multiuser MIMO uplink system. For either system, a mathematical optimization problem is formulated and tackled by deriving an iterative algorithm. Via numerical results, the mean squared error (MSE) performance of the digital and AirComp schemes is compared.