DEFL: A Novel Blockchain Fully-Orchestrated Federated Learning Framework

During the last few years, federated Learning (FL), an AI paradigm shift, has become an active research field, coming to allow the collaborative training of machine/deep learning models among different stakeholders without compromising privacy and security restrictions. Although researchers are attempting to support more FL-based architectures, federated learning systems (FLSs) are still facing challenges from various aspects such as privacy, consistency, and efficiency, especially in critical use cases where no stakeholder can be trusted for FL aggregation and validation. In this paper, we propose a novel Blockchain fully-orchestrated federated learning framework, we named DeFL, adapted for low or zero-trust configurations. We provide a thorough description of the proposed framework, its workflow and the different design components.