Scalable Federated Learning Simulations Using Virtual Client Engine in Flower
Thomás Borja Saltos, Daniel Alamillo, Anass Anhari, Ilker Demirkol · 2023
Federated Learning (FL) is a novel technology that has gained significant visibility among the scientific community and industry. In order to be deployed in real scenarios, Federated Learning must address some critical challenges. One key challenge that Federated Learning poses is system scalability, as it is foreseen that this technology will be deployed in IoT scenarios with billions of clients in the future. In addition, there are many ongoing efforts regarding algorithms and frameworks proposed to foster Federated Learning testing and research. In this work, we review a recent architecture of the open-source framework Flower, called Virtual Client Engine (VCE), which allows to efficiently run simulations with hundreds of users with acceptable accuracy and keeping the computational resource consumption low. We compare relevant classification metrics (accuracy, loss) to classic centralized machine learning approaches and Flower's previous architecture named Edge Client Engine (ECE) to show the improvement in the number of clients while keeping the accuracy high. For a 200 clients case, we achieve an accuracy of around 91% for the MNIST and 88% for the SPEECH evaluation setups, both cases using the VCE Flower architecture.