An Enhanced Privacy-Preserving Hierarchical Federated Learning Framework for IoV

J Luo, X Li, H Wang, D Lan, X Wu, L Zhou… - … Conference on Information …, 2023 - Springer
J Luo, X Li, H Wang, D Lan, X Wu, L Zhou, L Fang
International Conference on Information and Communications Security, 2023Springer
Abstract The intelligent Internet of Vehicles (IoV) can help alleviate road security issues.
However, increasing requirements for data privacy make it difficult for centralized machine
learning paradigms to collect sufficient training data, which hinders the development of
intelligent IoV. Federated Learning (FL) has emerged as a promising method to overcome
this gap. However, traditional FL may leak privacy when encountering attacks such as the
Membership Inference Attack. Existing approaches to address this issue either bring …
Abstract
The intelligent Internet of Vehicles (IoV) can help alleviate road security issues. However, increasing requirements for data privacy make it difficult for centralized machine learning paradigms to collect sufficient training data, which hinders the development of intelligent IoV. Federated Learning (FL) has emerged as a promising method to overcome this gap. However, traditional FL may leak privacy when encountering attacks such as the Membership Inference Attack. Existing approaches to address this issue either bring significant additional overhead or reduce the accuracy of FL, which are not suitable for the IoV.
Therefore, we present a novel hierarchical FL framework called EPHFL. It leverages the Diffie-Hellman algorithm and pseudorandom technology to enhance the privacy of FL while bringing little additional overhead and not reducing the accuracy. Its hierarchical architecture can effectively schedule devices in the IoV to accomplish FL and reduce the communication overhead of each device, dramatically improving our system’s scalability. Moreover, we design a method based on Blockchain and Distributed Hash Table to detect malicious tampering and offset its impact, further guaranteeing FL’s data integrity. Finally, we perform experiments to demonstrate the performance of EPHFL. The results show that our method does not reduce accuracy, and our computation overhead on the user side is much lower than the classic baseline.
Springer
Showing the best result for this search. See all results