Using Blockchain for Decentralized Artificial Intelligence with Data Privacy

Abstract

Artificial Intelligence (AI) solutions require model training, which traditionally needs local datasets to be uploaded to a centralized server. However, centralized learning usually creates issues in terms of data privacy, data control, and data security. To address these issues, decentralized AI, especially federated learning, is proposed to achieve no data sharing or data exchange across servers and decentralized devices. This paper proposes a new decentralized AI paradigm based on blockchain, which goes beyond the decentralized federated learning to require no model pooling or orchestrating on any central server. That is, we propose further decentralization by dismissing the central or-chestrator in federated learning through the usage of blockchain (e.g., Ethereum in this paper), which is inherently a decentralized, distributed, peer-to-peer and autonomous mechanism. We implemented a prototype to demonstrate the idea, in which models are trained on data contributors’ ends. Instead of moving models to a central place for combining (such as in federated learning), models are trained independently on data contributors’ sides, and flow to the decentralized blockchain for updating. In this process, smart contracts on Ethereum provide specifications required for model creation. In addition, data contributors are rewarded Ether coins for their contributions to train the model with their data. This also incentivizes data contributors to provide data for AI training while protecting their data privacy.