Deep Federated Fractional Scattering Network for Heterogeneous Edge Internet-of-Vehicle Fingerprinting: Theory and Implementation

Zhang, Tiantian ORCID: https://orcid.org/0000-0002-1817-7950, Xu, Dongyang ORCID: https://orcid.org/0000-0002-6401-0545, Ma, Jing ORCID: https://orcid.org/0000-0002-2791-7906, Bashir, Ali Kashif ORCID: https://orcid.org/0000-0003-2601-9327, Dabel, Maryam M. Al ORCID: https://orcid.org/0000-0003-4371-8939 and Feng, Hailin ORCID: https://orcid.org/0000-0003-2734-480X (2024) Deep Federated Fractional Scattering Network for Heterogeneous Edge Internet-of-Vehicle Fingerprinting: Theory and Implementation. IEEE Internet of Things Journal. ISSN 2372-2541

Preview

Accepted Version Available under License In Copyright. Download (2MB) | Preview

Abstract

With the rapid development of distributed edge intelligence (DEI) within Internet of vehicle (IoV) network, it is required to support heterogeneous rapid, reliable and lightweight authentication which prevents eavesdropping, tampering and replay attacks. Radio Frequency Fingerprinting (RFF), which leverages unique and tamper-proof hardware characteristics, is an emerging deep learning based physical layer technology poised to achieve excellent authentication within DEI enhanced heterogeneous IoV. However, centralized collection of critical datasets will bring severe privacy concerns as well as huge communication overheads towards resources-constrained IoV nodes. In this paper, we propose a deep federated fractional scattering fingerprinting network (FFSFNet) which amalgamates fractional wavelet scattering and federated learning to achieve excellent identification. Particularly, we first exploit fractional wavelet scattering to extract RFF characteristics from non-stationary waveform, eliminate redundancies and enhance interpretability. To improve the training efficiency and privacy protection capability, we design a novel federated framework, which not only completes distributed training, reduces overhead but also protects privacy. Furthermore, we conducted a comprehensive comparative analysis of different model quantization schemes and validated the proposed scheme with field programmable gate array (FPGA) accelerators. Experimental results demonstrate that the proposed FFSFNet can maintain excellent identification performance with only 5.08% of original samples. The model size and inference latency can be effectively improved by quantization with limited degradation. Moreover, the identification testing accuracy of FFSFNet can eventually converge to 99.4% with 0.64ms inference latency per sample.