OrderlessFL

Abstract

Industries produce a large amount of data that can improve Machine Learning models. However, due to privacy issues, the data cannot be shared. Several Federated Learning (FL) systems have been introduced as private alternatives without considering Byzantine actors. Also, these systems are affected by the gradient staleness problem. Several blockchain-based FL systems are introduced to address Byzantine actors, which rely on Proof-of-Work-based (PoW) protocols and suffer from their limitations. We introduce OrderlessFL, a safe permissioned blockchain-based FL system using flCRDT, a CRDT for concurrent ML training and mitigating gradient staleness.