Towards Optimizing Computational Costs of Federated Learning in Clouds

Federated Learning is a strategy where distributed training datasets are processed by several clients coordinated by a central server that keeps the global learning model. This approach is very attractive in the biomedical domain because it allows the use of private datasets from multiple institutions to train a model without the need of sharing the data. Moreover, Machine Learning approaches often require larger training samples than what can be afforded by a single institution. In this work, we are interested in analyzing the performance of a Tumor-Infiltrating Lymphocytes Classification problem when solved by a federated learning approach deployed in a commercial cloud. In the presented evaluation, we consider executions of a federated learning implementation on accelerated instance types of the AWS EC2, on on-demand and spot markets, while varying the number of clients. The obtained results showed the improvement of the accuracy and execution times when the number of clients increases. They also revealed that, although the spot instances suffered from revocations, their use could significantly reduce the financial costs compared to the on-demand one.