RePEL: Blockchain-Empowered Conditional Privacy-Preserving Encrypted Learning

Business organization performs its business activities within a headquarter-branch relationship, especially for multinationals, by establishing work places in accordance with business requirements. The data-driven decision making stimulates the importance of data analysis. Machine learning (ML), as a method of data analysis, automates analytical model building based on the idea that systems can train and learn from data to make decisions with minimal human intervention. For high accuracy, traditional ML algorithms make design tradeoffs, conceding privacy and reliability, and are thereby unable to satisfy strong security demands. To resolve design tensions, in this article, we propose RePEL, which harmonizes functional encryption (FE) and blockchain on top of encrypted learning. RePEL allows the headquarter to only share partial information about business data collected from branches while manages data transfer with the consensus mechanism in the blockchain, which works in a coordinated way for preserving conditional privacy and high reliability. We instantiate a RePEL design with the three-layer framework and formally reduce its security to a provably secure FE scheme. We further deploy Feel BC to implement a RePEL prototype system, so as to realize the performance evaluation. Experimental results show that RePEL, with the basic premise of privacy and reliability, can achieve high accuracy and reasonable throughputs.