Conceptual

Entropy-Based Data Selection with Partial Model Fine-Tuning for Efficient Federated Learning

FedFT-EDS, a federated-learning method that reduces the training workload on resource-constrained edge and IoT clients by combining two mechanisms: each client fine-tunes only part of its model instead of training the whole network, and an entropy-based data-selection rule keeps only the most informative local instances (about half the data) for each training round. The work shows that not all user data is equally beneficial for federated learning on every round; on CIFAR-10 and CIFAR-100 it improves the accuracy of the aggregated global model over the FedAvg and FedProx baselines while using only 50% of client data and up to three times less client training time.