Author(s): Esteves, Leonardo ; Portugal, David ; Peixoto, Paulo ; Falcão, Gabriel
Date: 2023
Persistent ID: https://hdl.handle.net/10316/113800
Origin: Estudo Geral - Universidade de Coimbra
Subject(s): federated learning (FL); federated averaging (FedAvg); federated SGD (FedSGD); unreliable participants; selective aggregation
Recent advances in artificial intelligence algorithms are leveraging massive amounts of data to optimize, refine, and improve existing solutions in critical areas such as healthcare, autonomous vehicles, robotics, social media, or human resources. The significant increase in the quantity of data generated each year makes it urgent to ensure the protection of sensitive information. Federated learning allows machine learning algorithms to be partially trained locally without sharing data, while ensuring the convergence of the model so that privacy and confidentiality are maintained. Federated learning shares similarities with distributed learning in that training is distributed in both paradigms. However, federated learning also decentralizes the data to maintain the confidentiality of the information. In this work, we explore this concept by using a federated architecture for a multimobile computing case study and focus our attention on the impact of unreliable participants and selective aggregation in the federated solution. Results with Android client participants are presented and discussed, illustrating the potential of the proposed approach for real-world applications.
