Federated Learning with Accurate Model Training and Low Communication Cost in Heterogeneous Scenarios

Federated learning (FL) is a distributed approach to train machine learning models without disclosing private data from participating clients to a central server. Nevertheless, FL performance depends on the data distribution, and the training struggles to converge when clients have distinct data distributions, increasing overall training time and the final model prediction error. This work proposes two strategies to reduce the impact of data heterogeneity in FL scenarios. Firstly, we propose a hierarchical client clustering system to mitigate the convergence obstacles of federated learning in non-Independent and Identically Distributed (IID) scenarios. The results show that our system has a better classification performance than FedAVG, increasing its accuracy by approximately 16% on non-IID scenarios. Furthermore, we improve our first proposal by implementing ATHENA-FL, a federated learning system that shares knowledge among different clusters. The proposed system also uses the one-versus-all model to train one binary detector for each class in the cluster. Thus, clients can compose complex models combining multiple detectors. ATHENA-FL mitigates data heterogeneity by maintaining the clustering step before training to mitigate data heterogeneity. Our results show that ATHENA-FL correctly identifies samples, achieving up to 10.9% higher accuracy than traditional training. Finally, ATHENA-FL achieves lower training communication costs than MobileNet architecture, reducing the number of transmitted bytes between 25% and 97% across evaluated scenarios.