Transfer-Informed Variational Autoencoder for Federated Learning on Imbalanced IoT data

Federated Learning (FL) enables privacy-preserving collaborative training but faces challenges like data heterogeneity and limited labeled data. This paper presents a new architecture combining class-informed variational autoencoders (CI-VAE) with transfer learning (TL) to improve FL performance under imbalanced data conditions. Fed by a pre-trained MobileNetV2 feature extractor, our method enhances image classification by establishing a shared latent space integrated with a linear classifier. Tested on the FEMNIST and IRDS datasets, it outperforms the Federated TL approach which doesn't relies on VAE, achieving 100% accuracy on FEMNIST and 91% on IRDS and showcasing its effectiveness for privacy-preserving, robust FL applications.