Variational AutoEncoders and Meta-Learning: Transforming Federated Learning in IoT Environments

This dissertation investigates the integration of Variational AutoEncoders and Meta-Learning in Federated Learning, particularly in the IoT domain, where data heterogeneity and limited samples pose significant challenges. The Class Informed-VAE (CI-VAE) is introduced, establishing a federated latent space that harmonizes data distributions of different clients into a unified distribution that also enables the generation of synthetic samples and labels. The proposed model attains performance on par with the conventional local and federated models across image datasets (FEMNIST) and sample-based datasets (HAR). Furthermore, the thesis explores the effectiveness of Meta-Learning (particularly the Reptile algorithm) in improving model performance with limited data, showcasing its potential in scenarios with few samples per class.