A Secure Federated Learning Framework using Blockchain and Differential Privacy

Abstract

Federated learning (FL) has considerably emerged as a promising solution to enhance user privacy and data security by enabling collaboratively multi-party model learning without exchanging confidential data. Nevertheless, most existing FL approaches still rely on a central server to obtain a global model by collecting all uploaded models from participants, which may lead to several threats from malicious participants and even expose participant privacy. Therefore, to tackle these problems, we proposed a secure FL framework by empowering blockchain to replace the centralized aggregator sever and utilize Differential Privacy (DP) to address various attacks, e.g., membership inference attacks, during the collaborative FL model training process. The proposed framework has been implemented through two scenarios, i.e., blockchain-based FL to form a decentralized system and DP-based FL to construct the randomized privacy protection using the IBM DP Library.