DQFed: A Federated Learning Strategy for Non-IID Data based on a Quality-Driven Perspective

The increasing amount of available smart objects produces a huge amount of data that can be successfully managed using Federated Learning (FL) approaches. FL is a distributed deep learning framework where several devices are trained with a local model on local data, and a central server aggregates them in a global model. These strategies are successfully used in several contexts ensuring privacy preservation and high effectiveness in data analysis. However, the FL strategies show variable accuracy in case of a huge presence of non-independent-and-identically-distributed (Non-IID) data. This challenge has been largely explored in the last years by researchers and developers who propose new FL strategies for non-IID data. This study introduces a novel FL approach, called DQFed, that aggregates the local models on the base of their weights computed according to a quality-driven model. The surrounding idea is that the performance of the general model can be improved by aggregating the clients' models giving higher importance to the clients that use high-quality data. The DQFed strategy is evaluated on five datasets (both non-IID and IID datasets are considered) obtained starting from a real dataset. The results show an improved F1-score compared to the considered baseline for both non-IID and IID data.