FedMcon: an adaptive aggregation method for federated learning via meta controller

• Proposes FedMcon, a meta-learning-based adaptive aggregation method that learns to aggregate heterogeneous local models in federated learning. • A learnable controller trained on a small proxy dataset replaces fixed aggregation rules, effectively addressing non-IID data distributions. • Achieves 19× communication speedup in a single FL setting while maintaining superior performance on extremely non-IID data. • Overcomes limitations of FedAvg's static linear combination weighting based solely on local data sizes.