On Adaptive Client/Miner Selection for Efficient Blockchain-Based Decentralized Federated Learning

This study presents a fast and accurate blockchain-based decentralized federated learning (BC-DFL) based on an adaptive client/miner selection algorithm. BC-DFL is a learning method that manages the machine learning models on a blockchain. Although the blockchain can improve the security of model sharing and realize reward management, its mining process extensively increases computation and communication load. The training accuracy could also be degraded if each client's data distribution follows non-independent and identically distributed (Non-IID) conditions. In the proposed method, a client selection allows parallel processing of local training and mining on the network, reducing round time. In addition, using a client selection algorithm based on the estimation of the label distribution, the accuracy degradation caused by Non-IID is suppressed. Numerical results demonstrate that the proposed method can improve training time and accuracy performances compared to related frameworks.