Privacy-preserving Blockchain-based Global Data Sharing for Federated Learning with Non-IID Data

Abstract

Federated learning is a popular privacy-enhanced distributed machine learning method that solves the problem of local data privacy by gathering the training results (such as model weights, gradients, etc.) instead of the raw data to generate a global model. But a practical problem it faces is the non-independent and identical distribution of data, which means the local data of each participant is highly inconsistent, both in terms of quantity and distribution. Moreover, there is a lack of research related to the efficiency and privacy issues in the pre-training process. Therefore, in this paper, we propose a novel solution that utilizes blockchain technology to realize small-scale global data sharing to assist the training progress. Simulation experiments verify that our method not only guarantees data security but also greatly improves performance in terms of training speed and accuracy.