Federated Learning for Heterogeneous and Inconsistent Data

PI: Anders Eklund, Linköping University
co-PI: Mikael Nilsson, Lund University

Federated learning (FL) promises to revolutionize AI by enabling collaborative training without sharing sensitive data (see figure). However, current FL methods fail catastrophically with real-world heterogeneous and inconsistent data, which cause model divergence and poor performance. This project will therefore develop novel global aggregation functions, novel methods for data harmonization, novel robust local models, as well as novel methods for fusion of multimodal data. Our vision is to be able to use FL for all kinds of sensitive heterogeneous and inconsistent data, which will have a large impact on several research fields.

Project number: G5