Robust Decentralized Federated Learning Using Collaborative Decisions

Abstract

Federated Learning (FL) has attracted a lot of attention in numerous applications due to recent data privacy regulations and increased awareness about data handling issues, combined with the ever-increasing big-data sizes. This paper proposes a server-less, robust FL training mechanism that allows any set of participating data-owners to train a neural network (NN) model collaboratively without the assistance of any central node and while being resilient to Byzantine attacks. The proposed approach makes use of a dual-way update mechanism to allow each node to take a model forwarding decision towards a global collaborative decision of isolating any malicious updates. The efficiency of the proposed approach in detecting cardiac irregularities is verified using simulation results conducted based on the Physikalisch-Technische Bundesanstalt Database electro-cardiogram (PTBDB ECG) dataset.