FedTwin: Blockchain-Enabled Adaptive Asynchronous Federated Learning for Digital Twin Networks

Abstract

The fast proliferation of digital twin (DT) establishes a direct connection between the physical entity and its deployed digital representation. As markets shift toward mass customization and new service delivery models, the digital representation has become more adaptive and agile by forming digital twin networks (DTNs). The DTN institutes a real-time single source of truth everywhere. However, there are several issues preventing DTNs from further application, including centralized processing, data falsification, privacy leakage, lack of incentive mechanism, and so on. To make DTN better meet the ever changing demands, we propose a novel block-chain-enabled adaptive asynchronous federated learning (FedTwin) paradigm for privacy-preserving and decentralized DTNs. We design Proof-of-Federalism (PoF), which is a tailor-made consensus algorithm for autonomous DTNs. In each DT's local training phase, generative adversarial network enhanced differential privacy is used to protect the privacy of local model parameters, while a modified Isolation Forest is deployed to filter out the falsified DTs. In the global aggregation phase, an improved Markov decision process is leveraged to select optimal DTs to achieve adaptive asynchronous aggregation while providing a rollback mechanism to redact the falsified global models. With this article, we aim to provide insights to forthcoming researchers and readers in this under-explored domain.