A Practical Evaluation of a Federated Learning Application in IoT Device and Cloud Environment
Guilherme Nunes Nasseh Barbosa, Diogo M. F. Mattos · Journal of the Brazilian Computer Society · 2025
Internet of Things (IoT) devices have grown exponentially in recent years, resulting in valuable data for machine learning applications. Traditionally, machine learning models require centralized data collection and processing, which is not feasible in the IoT landscape due to high density and growing data privacy concerns. Federated Learning is a trend in this scenario, as it allows collaborative training of models on IoT devices, distributed and without the need to share data. This paper examines a federated learning framework for IoT devices, employing a parameter server topology in a benign node environment without considering strategies for optimizing model performance. The evaluation is conducted in two distinct scenarios: (i) a testbed of IoT devices equipped with ARM processors and limited to 2GB of RAM, and (ii) a virtualized cloud environment with a mixture of resource-constrained virtual machines. The experiments use non-identically distributed (non-IID) datasets—MNIST for the IoT testbed and CIFAR-10 for the cloud environment—evaluated under various client configurations and aggregation strategies. In the IoT device scenario, the framework achieved an accuracy of up to 0.6 after ten rounds of global aggregation, while the cloud environment attained a maximum accuracy of 0.4. These results demonstrate the feasibility of applying FL in resource-constrained IoT environments, with scalability and accuracy influenced by the number of clients and local training epochs.