Enabling Flower for Federated Learning in Web Browsers in the Cloud-Edge-Client Continuum

Federated Learning (FL) is a solution that enables training a machine learning algorithm by distributing it across multiple devices (clients), which process a local model and then share the trained parameters to derive a global model. This has many advantages regarding computational distribution and privacy, as the clients' local datasets are never shared. Nevertheless, configuring and managing a cluster of heterogeneous devices can present several challenges. In this perspective, some frameworks, such as Flower, provide valuable tools to facilitate the construction of FL processes. Yet, infrastructure management, installation of dependencies, and network configuration are still required. We propose a solution that allows Flower to be seamlessly integrated with training within the Web browser, following the novel Cloud-Edge-Client Continuum paradigm. This enables ready-to-use devices, including laptops and smartphones, with no dependencies to install or configurations to make. The solution's effectiveness has been tested in heterogeneous environments exploiting Open Neural Network Exchange (ONNX) and Ten-sorFlow.js as deep learning technologies. Although browser-based training is generally slower than native execution, our evaluation shows no significant overhead when deployed in heterogeneous environments with resource-constrained edge devices.