FedDAFL: Federated Transfer Learning with Domain Adaptation for Frugally Labeled Datasets

While Federated Learning (FL) manages well data diversity in distrusted learning, it faces additional complexities in scenarios with "frugal labeling," where the client nodes host a mix of partially or fully unlabelled datasets. This paper introduces FedDAFL, a novel approach tailored for label-scarce federated settings, that enables learners with unlabelled datasets to leverage fully labelled teachers, thus incorporating domain adaptation techniques within a multi-teacher multi-learner framework. Through the integration of self-learning and semi-supervised methods, FedDAFL aims to enhance model accuracy using each client’s unlabelled data. Comparative analysis with established techniques like Federated Adversarial Domain Adaptation (FADA) and Federated Knowledge Alignment (FedKA), demonstrates FedDAFL’s superior accuracy across independent and identically or non-identically distributed (non-IID) datasets. Even if the clients possess single class data, FedDAFL adeptly tackles frugal labeling challenges in federated transfer learning, exhibiting competitive performance against existing methods. Results validated on Pneumonia chest X-ray datasets underscore FedDAFL’s adaptability in handling limited labels and non-IID setups, demonstrating its applications to various settings.