Reliable Collaborative Learning with Commensurate Incentive Schemes

Abstract

Collaborative learning techniques allow numerous clients conjointly to improve artificial intelligence models using their private datasets. The clients carry out the training locally and periodically exchanging gradient values through devices. Unlike conventional training approaches, the training data in the collaborative techniques are not revealed publicly. Regardless of privacy merits, clients are often less motivated to improve the model due to inadequate incentives procedural. In short, the resources owned are not maximally utilized. To tackle the issue, we design a collaborative learning model with a secure, fair, and immutable incentive mechanism by leveraging blockchain technology. Incentives are distributed proportionately to clients according to their respective contributions. We implement our incentive schemes on Ethereum. We also evaluate the performance of collaborative learning in a different setting. The results indicate that the design objectives are met.