PHiFL-TL: Personalized hierarchical federated learning using transfer learning

Federated Learning is a collaborative machine learning (ML) framework designed to train a globally shared model without accessing participants’ private data. However, due to the statistical heterogeneity in the participants’ data, federated learning faces significant challenges. This approach generates a similar output for all participants, without adapting the model to each individual. Consequently, the global model performs poorly on each participant's task. To mitigate these issues, personalized federated learning methods aim to reduce the negative effects caused by data heterogeneity. Previous personalized approaches have relied on a single central server. However, in federated learning based on a client-server architecture, the central server's workload becomes a bottleneck. In our paper, we propose a Personalized Hierarchical Federated Learning approach (PHiFL-TL). First, PHiFL-TL trains a global shared model using hierarchical federated learning. Next, it constructs relatively personalized models through transfer learning. We demonstrate the effectiveness of PHiFL-TL on non-identical and independent data partitions from MNIST and FEMNIST datasets.