Privacy-preserving and Personalized Federated Machine Learning for Medical Data

The federated learning setting is prone to suffering from non-identically distributed data across participating agents. This gives rise to the task of model personalization, where agents collaborate to train several different machine learning models instead of training only one global model. The aim of model personalization is to minimize the sum of the generalization error incurred from training on small data sets, and the transfer error incurred from applying a globally-trained model to a specific local distribution. In this report, two novel approaches to personalized cross-silo federated learning are introduced and discussed from a theoretical perspective: the adapted Ndoye factor, and the Weight Erosion aggregation scheme. The latter is implemented and compared to two baseline aggregation schemes in two case studies: training a diagnostic model on real-world medical data, and predicting the survival of passengers on the publicly available Titanic data set. The models trained using the Weight Erosion aggregation scheme are compared to those trained using the baseline aggregation schemes, both in terms of their classification accuracy on the local test set and in terms of the learned parameters. We demonstrate that the novel Weight Erosion scheme can outperform both baseline aggregation schemes for some specific tasks.