Semantic-Based Dimensionality Reduction in Federated Learning Approaches

To overcome one of the main limitations of federated learning, that is the non-negligible communication overhead between the clients and server, the present work proposes a novel federated scheme based on principles envisaged by semantic communications. The proposed semantic-based dimensionality reduction algorithm is employed to reduce the data exchanges by more than one order of magnitude and negligible performance loss. The effectiveness of the proposed approach is validated through a classification scenario leveraging transfer learning.