A Reputation-Based Learning Nodes Selection Algorithm for Decentralized Federated Learning

Due to the increasing risk of data security, distributed learning model based on real-world data analytics has attracted more attention, and it has been applied in a variety of areas ranging from medical screening to agriculture, industry, finance, and defense science. Generally, participants provide their own private datasets to efficiently train the distributed models on real-world data, which inevitably leads to privacy and security concerns. Without uploading raw training data, Federated Learning enables large-amount nodes to train a distributed model and preserves security and privacy of user sensitive information. However, Federated Learning is limited by expensive computational costs during collective parameter server aggregation. Moreover, malicious nodes among computing nodes interfere model training to some extent and further cause the leakage of data privacy. To address the above-mentioned problems, we propose a novel Decentralized Federated Learning by integrating blockchain and federation learning for efficient node selection and communication. Based on the proposed model, a reputation-based learning nodes selection algorithm is presented to measure the probability of honest participation of distributed nodes. The simulation results demonstrate that our RBLNS is capable of improving the training result significantly and decreasing convergence time.