Quantization-Based Federated Learning in User-Centric Cell-Free Massive MIMO

User-centric cell-free massive multiple-input multiple-output (CFmMIMO) is a fundamental architecture for future mobile networks, characterized by its capability to offer macro diversity, enhanced energy efficiency, interference management, and reliable connectivity. Concurrently, federated learning (FL) gains significant attention as a distributed machine learning paradigm. The large number of model parameters and the massive connectivity lead to a significant communication overhead of FL. Recent efforts on FL are dedicated to its deployment in CFmMIMO but do not take communication efficiency into account. In this paper, we introduce compression and specifically investigate quantization-based FL (QBFL) over user-centric CFmMIMO. In view of uplink transmission, we first apply a low-precision stochastic quantizer to model updates for reducing the amount of data transmission. We then delve into the scenario of partial user participation within QBFL over user-centric CFmMIMO. The convergence of the QBFL is evaluated via both theoretical analysis and experimental simulations. Moreover, numerical results show that QBFL guarantees learning performance and reduces communication overhead.