Blockchain-Based Participant Selection for Federated Learning

Abstract

Federated Learning (FL) advocates training a global model using a large-scale of distributed devices such that the collaborative model training can be benefited from the rich local datasets, while preserving the privacy of local training dataset in each participant. This is because only the training results, i.e., the updated model parameters and model weights are needed to report to the FL server for aggregation in each round of FL-training. However, during the model transmission of the original FL protocol, there is no security guarantee towards the training results. Thus, every step during model uploading phase can be attacked by malicious attackers. To address such threat, we propose a new federated learning architecture by taking the advantages of blockchain into account. The proposed architecture includes two-phase design. The first phase is a numerical evaluation, which can prevent the malicious devices from being selected. For the second phase, we devise a participant-selection algorithm that enables the FL server to select the appropriate group of devices for each round of FL-training. We believe that our study can shed new light on the joint research of blockchain and federated learning.