SecurePrivChain: A decentralized framework for securing the global model using cryptography

Kahkishan Sanam

doi:10.1016/j.future.2022.12.032

2023-05-01

PRE

Kahkishan Sanam

Saif Ur Rehman Malik *

Tehsin Kanwal

Zain Ul Islam Adil

DOI：10.1016/j.future.2022.12.032

原文链接

原文求助

摘要

En 中文

VANETS (IoVs), banks, and healthcare records are the sensitive information of vehicles, clients, and patients that is stored and maintained electronically, which has historically been a popular target for privacy leakage. To address growing issues about privacy leakage in these scenarios, federated learning has been frequently accepted as a privacy-preserving framework that enables multiple users to collectively develop a global model. To preserve the privacy of the training data, the server and participants share only model parameters. However, since participants have access to modified model parameters, this solution is still prone to poisoning and label-flipping attacks. Recent research showed that the federated learning framework is highly susceptible to attacks, with poisoning and label flipping being the most powerful, as well as secluded attacks, in which attackers can poison the parameters of the global model to gain access to the clients' confidential information or degrade the global model's performance. To address these attacks, a decentralized framework for securing the global models in federated learning (SecurePrivChain) is proposed in this study. A permissioned private Ethereum blockchain and encryption mechanisms were adopted in the framework to hide global parameters and preserve the privacy of a client's data as well as achieve client's authentication. Finally, the proposed model is evaluated using metrics such as computed cost, transaction latency, loss, and cost analysis based on Ethereum gas consumption. The performance of the scheme was compared with other benchmark models which are based on different encryption mechanisms. These included RSA and algamal using publicly available MNIST, banks, vehicles, and hospital datasets. Numerical results show that the proposed framework has good performance. A formal security analysis is done to show how the system can meet its goal of protecting privacy.(c) 2022 Elsevier B.V. All rights reserved.

Keyword:

Horizontal federated learning

Private Ethereum

ECC

RSA

Algamal

Internet of vehicles

Cryptography

Global model security