Reproduce “BAFFLE : Blockchain Based Aggregator Free Federated Learning” on OpenResearch

About this paper

Paritosh Ramanan, Kiyoshi Nakayama

Researchers from Georgia Institute of Technology and TieSet Inc. developed BAFFLE, an aggregator-free Federated Learning environment built on blockchain technology. This system leverages Ethereum Smart Contracts to coordinate machine learning tasks, addressing limitations of centralized FL by enabling decentralized model updates and aggregation through mechanisms like model chunking and a score-and-bid strategy. BAFFLE achieves model quality comparable to centralized FL while demonstrating reduced blockchain gas costs and improved scalability in a deep reinforcement learning application for taxi revenue optimization.

Key insights

• BAFFLE's design with model chunking and a score-and-bid strategy effectively mitigates blockchain's computational and transactional constraints, enabling efficient on-chain machine learning.
• The learning process of BAFFLE is theoretically equivalent to classical Federated Learning, ensuring comparable model convergence and accuracy despite the absence of a central aggregator.
• Leveraging blockchain for an aggregator-free framework significantly lowers the barrier to entry for small organizations and enhances trustworthiness by decentralizing coordination and updates.

Summary by alphaXiv.