Optimizing Radio Frequency Fingerprinting for Device Classification: a Study Towards Lightweight Dl Models

Abstract

As the Internet of Things (IoT) permeates diverse application domains, ensuring the security of wireless networks has become increasingly critical. However, the constraints of resource-limited IoT devices render complex encryption impractical. Consequently, Radio Frequency Fingerprinting (RFF) has emerged as a promising avenue, leveraging unique device characteristics resulting from manufacturing nonlinearities. RFF enhances physical layer security by enabling device classification and authentication at IoT gateways. While deep learning (DL) aided RFF systems offer exceptional classification accuracy, their deployment on edge devices remains challenging to this end. Accordingly, there is a gap in the literature for efficient model exploration and implementation. This study proposes a lightweight Convolutional Neural Network (CNN) model using 1D convolutional filters to reduce inference latency. The model was applied to an open-source dataset comprising 30 LoRa devices. An evaluation was conducted to compare classification accuracy and inference latency using Short Time Fourier Transform (STFT) and Fast Fourier Transform (FFT) for preprocessing. Additionally, the performance of the proposed model was compared against a CNN model utilizing 2D convolutional filters. The model exhibited a significant reduction in inference latency with miniscule degradation in classification accuracy, addressing the identified gap, and propelling the academic discourse towards RFF for edge devices.