Federated learning is a machine learning approach that allows deep learning models to be trained on local servers and aggregated on a global server by uploading the parameters and keeping all data on the local servers. Federated-Averaging, which is the earliest and most popular federated learning algorithm, is compared with new alternatives that add improvements, Federated Dynamic-regularization and Federated- Ensemble. In this work, a comparison between these three supervised algorithms has been performed in a breast cancer classification problem using ultrasound images. Usually, medical organizations are unwilling to share their data with external servers due to data privacy constraints and, furthermore, due to the lack of data from each institution, usually more than one institution has to be involved in the training process.
Therefore, this is a perfect field where federated learning can be applied. In this comparison, the performance of these methods have been analyzed at both global server and client level, so it has been possible to carry out an assessment of the impact of the number of clients and the amount of data used by each client on each algorithm. It has been observed that the new FedENS algorithm yields to higher accuracy at global level, more stable results at client level and is less affected by client distribution. Keywords— Ensemble, Federated Learning, Deep Learning, Breast cancer classification, FedENS
Academic Information
Attached Resources
