An Analysis about Federated Learning in Low-Powerful Devices
Daniel Flores-Martin, Jaime Galán–Jiménez, Javier Berrocal, Juan M. Murillo · 2023
Federated Learning allows us to train Machine Learning models in a distributed way. This improves users' security and privacy and allows the computational load to be distributed. One of the advantages is the application of these models on low-powerful devices, bringing the processing closer to the devices that generate them, as is pursued with Edge Computing for the continuum. However, there is a need to conduct more extensive studies on the limitations of these devices such as the lack of processing power, limited memory available, or training time. Current works highlight these constraints and emphasize real experiments that evaluate the performance and limitations of these models. Therefore, this paper presents an analysis that evaluates the training and execution requirements, limitations, and resource consumption by running a Federated Learning algorithm proposed by Google and using the dataset Fashion-MNIST on different devices. This allows us to determine the requirements when selecting devices to be included in the federations.