Power Allocation for Efficient Decentralized Federated Learning over Cell-Free Massive MIMO

Cell-free massive multiple-input multiple-output (CFmMIMO) and federated learning (FL) are key enablers for distributed intelligence in future sixth-generation (6G) networks. However, the uplink transmission of high-dimensional local models in FL over wireless networks incurs significant communication overhead, which limits training efficiency under practical latency and energy constraints. In this paper, we study decentralized federated learning (DFL) over CFmMIMO networks, where interconnected access points collaboratively aggregate models and focus on uplink power allocation to enable efficient model training. We formulate a joint uplink latency and energy optimization problem by incorporating the achievable uplink rate. Based on the derived analytical expressions, an efficient coordinate descent-based power allocation algorithm is developed to balance the latency-energy trade-off in DFL global training. Simulation results show that our proposed method improves the final model accuracy by up to 32% compared with benchmark schemes under various resource budget constraints. These results demonstrate the effectiveness of latency and energy considered power allocation for efficient DFL over CFmMIMO networks.