Blockchain-based distributed federated learning using proof of accuracy consensus

Abstract This paper explores integrating federated learning (FL) and blockchain tech- nology, two burgeoning fields in information technology. Despite their growing popularity, both domains face significant challenges. In federated learning, the primary concern is safeguarding the integrity of the general model against client- induced compromises. Blockchain technology grapples with the need for a green mining approach through an energy-efficient consensus protocol. Our study lever- ages the strengths of each platform to mitigate the weaknesses of the other. We introduce an innovative blockchain-based FL model that eliminates the need for a central aggregator. Utilizing a green mining consensus algorithm named Proof of Accuracy (PoA), we create a competitive environment among nodes, fostering the creation of superior models. This approach ensures data integrity and model validation through a community-based consensus, resulting in a fully distributed system. This system enhances FL’s security and scalability and addresses vulner- abilities like malicious aggregators and scalability issues. Through experimental evaluations on the MNIST dataset with 20 miners, on one hand, our method enhances model accuracy to nearly 99% only after 10 blocks which is a higher point compared to FL and central learning. On the other hand, replacing Proof of Work (PoW) with PoA reduces energy consumption by nearly 30%. More- over, blockchain attacks appeared to be inapplicable, or resolvable after 6 blocks like fork attacks. After all, the introduced incentivizing mechanism lets malicious nodes get nearly zero rewards and allocates main rewards to honest nodes which is coherent with their efforts to present a superior model.