A Blockchain-Enabled Decentralized Gossip Federated Learning Framework

Abstract

Federated learning (FL) has undergone substantial research and has been used in numerous real-world solutions over the past few years. It shows promising results in addressing the data security and privacy issues present in the traditional centralized machine learning approach. Even though federated learning makes certain of data privacy for each contributing user, the global model and the data are still vulnerable to attacks by compromised clients and servers. Additionally, in the settings for non-independent identical data (Non-IID), federated learning performs significantly less compared to the standard centralized learning mode. To address both the security and performance issues, this paper proposes a blockchain-enabled gossip federated learning framework (BGFL). BGFL replaces the central server with a blockchain-enabled system for global model storage and exchange. Also, to achieve faster training convergence, clients communicate with each other based on a gossip training approach. Then, to evaluate the performance we perform experiments using MNIST and CIFAR datasets in Non-IID settings. The performance and effectiveness of the BGFL framework is demonstrated by the experimental results.