AdaFed: Performance-based Adaptive Federated Learning

Abstract

Federated Learning is a distributed and privacy-preserving machine learning technique that allows local clients to learn a model without sharing their own data by coordinating with a global server. In this work, we present the Adaptive Federated Learning (AdaFed) algorithm, which aims at improving the training performance of deep neural networks in Federated Learning settings by: (i) dynamically weighting the local models in the model averaging procedure; (ii) by adapting the loss function used by the federation at every communication round. We discuss the specialisation of AdaFed for both classification and regression tasks, providing several validation examples. Due to its adaptive design, the AdaFed algorithm showed a robust behaviour against unbalanced data distributions and adversarial clients.