In the era of artificial intelligence, trained neural network models have become new products of the information age. Most of machine learning strategies currently used to train neural networks are supervised learning, and thus, training data with labels become new intellectual capital (IC). Due to commercial confidentiality, data cannot be shared directly among information companies, which in turn prevents them from integrating resources to train better models. We need a framework that encrypts neural network models trained on the data and provides certain model exchange rewards that can be used to incentivize data sharing and to protect intellectual property (IP) and privacy of intellectual capital. Currently, federated learning provides a framework to train neural networks without compromising privacy, while block chain–based trading systems can attract other participants through a reward mechanism set by smart contracts. In this paper, we propose a block chain–based federated learning algorithm that enables reliable data sharing while protecting data from leakage, and design smart contracts based on the incentive mechanism of Shapley Values to reward data providers. We design a platform for managing IC by combining federated learning and block chain called Federated Learning Intellectual Capital Platform (FedLICP).

在人工智能时代，经过训练的神经网络模型已经成为信息时代的新产品。目前用于训练神经网络的机器学习策略大多是监督学习，因此，带有标签的训练数据成为新的智力资本（IC）。由于商业机密，信息公司之间不能直接共享数据，这反过来又阻碍了他们整合资源来训练更好的模型。我们需要一个框架来加密在数据上训练的神经网络模型，并提供某些模型交换奖励，可用于激励数据共享和保护知识产权 (IP) 和智力资本的隐私。目前，联邦学习提供了一个框架来训练神经网络而不损害隐私，而基于区块链的交易系统可以通过智能合约设置的奖励机制吸引其他参与者。在本文中，我们提出了一种基于区块链的联邦学习算法，在保护数据不被泄露的同时实现可靠的数据共享，并基于 Shapley Values 的激励机制设计智能合约来奖励数据提供者。我们通过结合联邦学习和区块链设计了一个管理 IC 的平台，称为联邦学习智力资本平台 (FedLICP)。