Privacy-preserving blockchain-enabled federated learning for B5G-Driven edge computing

【Author】 Wan, Yichen; Qu, Youyang; Gao, Longxiang; Xiang, Yong

【Source】COMPUTER NETWORKS

【影响因子】5.493

【Abstract】The arrival of the fifth-generation technology standard for broadband cellular networks (5G) and beyond 5G networks (B5G) rises the speed and robustness ceiling of communicating networks and thereby empowers the rapid popularization of edge computing. Consequently, B5G-Driven edge computing allows a growing volume of data to be collected from and transmitted among pervasive edge devices for big data analytics. The collected big data becomes the driving force of artificial intelligence (AI) by training high-quality machine learning (ML) models, which is followed by severe individual privacy leakage. Federated learning(FL) is then proposed to achieve privacy-preserving machine learning by avoiding the exchange of raw data. Unfortunately, several major issues remain outstanding. Centralized processing costs significant communication resources between cloud and edge while data falsification problems persist. In addition, the private data may be reconstructed by malicious participants by exploiting the context of model parameters in FL. To solve the identified problems, we propose to integrate blockchain-enabled FL with Wasserstein generative adversarial network (WGAN) enabled differential privacy (DP) to protect the model parameters of edge devices in B5G networks. Blockchain enables decentralized FL to reduce communication costs between cloud and edge while alleviating the data falsification issues, and it also provides an incentive mechanism to alleviate the data island issue in B5G-Driven edge computing. WGAN is used to generate controllable random noise complying with DP requirements, which is then injected to model parameters. WGAN-enabled DP is able to achieve an optimized trade-off between differential privacy protection and improved data utility of model parameters. Time delay analysis is conducted to show the efficiency of the proposed model. Extensive evaluation results from simulations demonstrate superior performances from aspects of convergence efficiency, accuracy, and data utility.

【Keywords】Blockchain; Federated learning; Differential privacy protection; Wasserstein generative adversarial nets

【发表时间】2022 FEB 26

【收录时间】2022-01-18

【文献类型】期刊

【主题类别】

区块链技术--