• clustering
• data privacy
• decision trees
• federated learning
• personalization

This thesis examines the use of the ID3 decision tree for Federated Learning experiments to guarantee interpretability and address the limitations of previous deep learning-based FL approaches. The study examines three levels of data distribution among clients: fully IID, partially non-IID, and completely non-IID, using the N-BaIoT dataset related to IoT network attack detection. Two FL methods were explored: the first approach selects the most representative model among local decision trees, proving to be the best choice for IID data distributions due to its performance and computational efficiency. However, it is not able to generalize in non-IID scenarios due to variations in data distribution across devices. The second approach employs IBM’s Federated ID3 decision tree framework, where clients compute class distribution statistics for feature values, and the server aggregates these to determine optimal splits by minimizing entropy. This method shows higher performance in non-IID settings. To further improve performance, personalization was implemented by clustering clients based on prediction similarity using hierarchical clustering, optimizing the number of clusters through the silhouette score. Within each cluster, the same FL approaches were applied separately, forming distinct federations.