Abstract
Self-supervised learning in federated learning paradigm has been gaining alot of interest both in industry and research due to the collaborative learningcapability on unlabeled yet isolated data. However, self-supervised basedfederated learning strategies suffer from performance degradation due to labelscarcity and diverse data distributions, i.e., data heterogeneity. In thispaper, we propose the SelfFed framework for Internet of Medical Things (IoMT).Our proposed SelfFed framework works in two phases. The first phase is thepre-training paradigm that performs augmentive modeling using Swin Transformerbased encoder in a decentralized manner. The first phase of SelfFed frameworkhelps to overcome the data heterogeneity issue. The second phase is thefine-tuning paradigm that introduces contrastive network and a novelaggregation strategy that is trained on limited labeled data for a target taskin a decentralized manner. This fine-tuning stage overcomes the label scarcityproblem. We perform our experimental analysis on publicly available medicalimaging datasets and show that our proposed SelfFed framework performs betterwhen compared to existing baselines concerning non-independent and identicallydistributed (IID) data and label scarcity. Our method achieves a maximumimprovement of 8.8% and 4.1% on Retina and COVID-FL datasets on non-IIDdataset. Further, our proposed method outperforms existing baselines even whentrained on a few (10%) labeled instances.