Collaborative Governance: Blockchain-Based Federated Learning for Construction Safety Service
针对施工安全服务中深度学习模型训练面临的数据安全与隐私问题，提出基于区块链的联邦学习框架，通过共享模型权重而非原始数据，并设计区块链安全共享策略，提升模型训练效果，帮助管理者理解项目安全绩效并做出管理决策。
Deep learning (DL) models are increasingly used to identify unsafe activities for the construction safety service (CSS). However, two typical issues threaten DL training process performance: poor security and privacy in data sharing. To address these problems, a blockchain-based federated learning (BCFL) framework is proposed from a collaborative governance perspective to obtain optimistic DL models for CSS. Two special works of this BCFL framework are that 1) it develops the federated learning empowered privacy-preserving data sharing with the principle of sharing data model weights instead of raw data (especially for one DL application for CSS, the Fed-YOLOv4 model for workers’ unsafe behavior identification task is developed) and 2) it explores two blockchain-based secure model sharing strategies that involve blockchain–interplanetary file system combination and model training contribution computing. Then, the smart contracts are further developed with the strategies above to streamline the workflow of federated DL model training. Finally, we apply the proposed framework to the practical subway project. The results demonstrate that the proposed framework can improve the DL model training and acquire global DL models for CSS with good accuracy. Our findings indicate that great data security and privacy ensured by introducing blockchain and federated learning can optimize data sharing and then DL models can be improved. Moreover, this study provides managers with a collaborative governance perspective on how DL models can be improved and further applied to CSS. This enables managers to quickly understand project safety performance and make timely managerial decisions for construction management.