Towards Collaborative Privacy-preserving Machine Learning on Private Blockchain

Approaches combining machine learning and blockchain technologies usually propose transferring the majority of the machine learning process to smart contracts, without considering implications on how it can impact the runtime complexity and flexibility of the system. In this study, we propose a method to enable distributed machine learning using blockchain without disclosing sensitive training data, or data stored on blockchain. The proposed method separates model training, model deployment, and model usage stages, allowing increased privacy. To demonstrate the possibility to apply the proposed method in practise, we developed a prototype that employs smart contracts, local blockchain Oracles, and distributed application that allows user to interact with blockchain network via graphical user interface. The prototype system was deployed into the Hyperledger Fabric environment and experimentally tested combining multiple machine learning models into a Shapley-weighted ensemble. We have shown that our method can be implemented using existing blockchain technologies and can be used to deploy data and machine learning models. Future developments would consider the network-based ensemble as a teacher for transferring knowledge to achieve privacy-preserving machine learning.