Model-Oriented Orchestration for Federated Learning: Personalization Through Intelligent Client Matching

The proliferation of connected devices and the exponential growth of generated data have revealed several limitations of classical cloud-centric architectures, notably in latency, bandwidth saturation, and data privacy. In response to these challenges, fog and edge computing paradigms have emerged, promoting decentralized processing by bringing computation and storage capabilities closer to data sources. Federated learning is a distributed paradigm for training machine learning models, to preserve data privacy and reduce communication costs. Among the widely adopted strategies are federated stochastic gradient descent, which aggregates client-side gradients, and federated averaging, which computes an average of local model parameters. Nevertheless, these approaches often overlook disparities in data quality and the varying contributions of participating clients, which may degrade model performance, particularly in non-iid or heterogeneous settings. In this paper, we introduce a novel aggregation framework for federated learning, leveraging an orchestrator model to identify the client whose data distribution most closely aligns with the given input, rather than applying uniform aggregation across all client models, as done in FedAvg. In our methodology, each client employed a distinct machine learning algorithm trained on three categories of datasets: first related to human activity recognition, second the other comprising MRI data and MIMIC-III dataset. The proposed approach achieved high accuracies of 99.90% on the UCI-HAR dataset, 99.13% on the MRI dataset and 99.55% on the MIMIC-III dataset, and also, we compared the proposed approach with other well-known personalized federation methods such as FedPro, FedPer, pFedHN and Fed-Clustering.