Blockchain-Based Decentralized Federated Learning With On-Chain Model Aggregation and Incentive Mechanism for Industrial IoT

• Published in: IEEE open journal of the Communications Society Vol. 5; pp. 6420 - 6429
• Main Authors: Yang, Qing, Xu, Wei, Wang, Taotao, Wang, Hao, Wu, Xiaoxiao, Cao, Bin, Zhang, Shengli
• Format: Journal Article
• Language: English
• Published: New York IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Blockchain; Blockchains; Computational modeling; Computer architecture; Damage detection; Data models; Decentralized federated learning; Federated learning; incentive mechanism; Industrial applications; Industrial Internet of Things; industrial IoT; Machine learning; Privacy; Servers; Smart manufacturing; Training; Training data
• ISSN: 2644-125X; 2644-125X
• DOI: 10.1109/OJCOMS.2024.3471621

Federated learning (FL) is an emerging machine learning paradigm that enables the participants to train a global model without sharing the training data. Recently, FL has been widely deployed in industrial IoT scenarios because of its data privacy and scalability. However, the current FL architecture relies on a central server to orchestrate the FL process, thus incurring a risk of privacy leakage and single-point failure. To address this issue, we propose a fully decentralized FL architecture based on blockchain technology. Unlike existing blockchain-based FL systems that use blockchain for coordination or storage, we use blockchain as a trustable computing platform for model aggregation. Furthermore, we model the interaction between the FL task publisher and participants as a Stackelberg game and design a rewarding mechanism to incentivize participants to contribute to the FL task. We build a prototype system of the proposed decentralized FL architecture and implement an FL-based damaged package detection application to evaluate the proposed approach. Experimental results show that the blockchain-based decentralized FL is feasible in a practical industrial IoT scenario, and the incentive mechanism performs well with real application data.