BLOCKCHAIN-ASSISTED PRIVACY AND SECURITY ENHANCEMENT IN FEDERATED LEARNING

Abstract

Clients can securely share gradients computed on their local data with the server using Federated Learning (FL), removing the need for them to directly expose their sensitive local datasets. During the process of model aggregation in traditional FL, the server might make use of its dominant position to infer sensitive information from the shared gradients of the clients. During model training, malicious clients may also submit forged and malicious gradients. Such behavior not only compromises the integrity of the global model, but also diminishes the usability and reliability of trained models. To effectively address such privacy and security attack issues, this work proposes a Blockchainbased Privacy-preserving and Secure Federated Learning (BPS-FL) scheme, which employs the threshold homomorphic encryption to protect the local gradients of clients. We develop a Byzantine-robust aggregation protocol for BPS-FL to implement cipher-text level secure model aggregation in order to ward off malicious gradient attacks. Furthermore, the immutability and traceability of the data are guaranteed by our use of a blockchain as the fundamental distributed architecture for recording all learning processes. Our extensive security analysis and numerical evaluation demonstrate that BPS-FL satisfies the privacy requirements and can effectively defend against poisoning attacks.