Optimization of federated learning topology in Internet of Things ecosystems

Concerns about preserving the privacy of data used in Machine Learning have been rising steadily for the last couple of years. Similarly, a big focus in current research is decreasing the amount of energy needed for training a state-of-the-art model due to the significant carbon footprint of current data centers. Meanwhile, the number of smart appliances that have substantial computational capabilities and connections to robust Internet of Things infrastructures, and yet are not meaningfully engaging these resources is only growing. Federated learning, a paradigm that proposes training the model on a federation of clients with a central server aggregating the updates, has the chance to mitigate both of these concerns by leveraging the full capabilities of IoT devices. However, for it to happen, stable solutions providing a reliable, efficient and convergent Internet of Things implementation of federated learning have to be developed. This study will focus on testing and evaluating the performance of various topologies against specific limitations of edge computing environments. To do so, a set of system components providing the functionalities of federated learning clients and server was implemented. It was then used for testing and analysis of various topologies