Quantum-Enhanced Communication-Efficient Federated Learning With Dynamic Power Allocation for Massive MIMO Systems

Abstract

With the widespread deployment of Massive multiple-input multiple-output (Massive MIMO) in 5G/6G networks, the integration of federated learning (FL) faces critical challenges such as limited uplink (UL) bandwidth, wireless channel impairments, and inefficient resource allocation. This paper proposes a quantum-enhanced communication-efficient federated learning (QE-CEFL) framework for supervised binary classification to tackle these bottlenecks. First, we adopt variational quantum circuits (VQCs) to replace conventional multi-layer perceptrons (MLPs), reducing the UL communication payload by 98% via extreme parameter compression. Second, we design a dynamic power allocation strategy that schedules users according to data significance, gradient importance, and channel gain, thus maintaining robust learning performance under non-independent and identically distributed (non-IID) data and improving energy efficiency. Third, we develop a partial priority aggregation mechanism to balance data volume and update importance for enhanced convergence stability. In addition, by exploiting Massive MIMO channel symmetry, we investigate the feasibility of time-division duplexing (TDD) reciprocity-enabled UL-assisted downlink (DL) broadcasting. While classical MLPs offer higher parameter capacity, the proposed framework achieves superior communication efficiency and energy savings via model compression, dynamic power allocation, and priority aggregation. Experimental results validate that QE-CEFL achieves a minimum test loss of 0.1114 and a test accuracy of 96.35% in a representative trial (25-trial mean accuracy: $95.8\% \pm 1.7\%$ ) with reduced communication overhead. The extreme model compression also effectively mitigates training oscillations incurred by MLP overparameterization. Energy efficiency is examined from two complementary perspectives: under a fixed convergence speed, the proposed framework achieves a 15-fold energy reduction (ranging approximately from $13\times $ to $17\times $ ) compared to the baseline, while under an equal-power constraint, it consumes only 30% of the baseline’s total energy to reach 90% accuracy due to faster convergence. Scalability experiments further verify that larger antenna arrays accelerate convergence and elevate final classification accuracy, demonstrating the inherent channel hardening gain of large-scale Massive MIMO systems.