Communication efficiency is central to over-the-air federated learning (OTA-FL), especially when edge devices operate over bandwidth-limited networks. OTA-FL exploits wireless multiple-access superposition to aggregate uplink updates, but the downlink model distribution and redundant late training rounds can still dominate the end-to-end communication cost. We present FedRTAS, a practical OTA-FL communication optimization framework that combines Replay-Based Training (RBT) and Trend-Aware Stopping (TAS). RBT caches recent server-side aggregated updates and reuses their average in server-only replay rounds, skipping both uplink and downlink communication during those rounds. It also applies FP16 quantization and gzip compression to the downlink global model payload. TAS monitors a validation accuracy trajectory with a polynomial-smoothed EMA slope and stops training when the trajectory reaches a plateau. Experiments on three medical image datasets and three general image classification datasets show that FedRTAS improves accuracy per transmitted byte under fixed communication budgets. Module-level ablations examine the effects of RBT and TAS, and the Non-IID Sensitivity Analysis shows how replay behaves as client heterogeneity increases. The results support FedRTAS as a modular OTA-FL strategy for resource-constrained image classification, including privacy-preserving medical deployments.
Gao et al. (Fri,) studied this question.