With the rapid advancement of deep learning techniques, numerous neural networks have been successfully developed for radio frequency (RF) fingerprinting identification. In this work, we propose a lightweight yet reliable neural network framework featuring a 9-layer architecture based on the long short-term memory (LSTM) strategy, designed for efficient open-set fingerprinting identification. The simulated beacon frames model real-world propagation effects by incorporating random modulation, power amplifier nonlinearity, multi-path fading, inherent radio noise, and additive channel noise. We extensively evaluate the identification accuracy and efficiency of our LSTM network identification against well-known deep learning models such as ResNet (144 layers) and GoogleNet (177 layers). The evaluation covers a wide range of parameters, including transmitter variability (s), number of transmitters (N), frames per transmitter (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(FPT\)</EquationSource> </InlineEquation>) and signal-to-noise ratio (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(SNR\)</EquationSource> </InlineEquation>). Our results show that the LSTM network maintains an accuracy of more than <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(96\%\)</EquationSource> </InlineEquation> in <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(SNR \ge {20}\)</EquationSource> </InlineEquation> with <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(FPT =200\)</EquationSource> </InlineEquation>, even with up to <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(N=1000\)</EquationSource> </InlineEquation> transmitters. At lower values (<InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(FPT \le {100}\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(6\le SNR \le {15}\)</EquationSource> </InlineEquation> dB), our LSTM network outperforms GoogleNet and matches ResNet in accuracy. Furthermore, it achieves a training acceleration of up to <InlineEquation ID="IEq9"> <EquationSource Format="TEX">\(116.7\times\)</EquationSource> </InlineEquation> for <InlineEquation ID="IEq10"> <EquationSource Format="TEX">\(N=100\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq11"> <EquationSource Format="TEX">\(FPT =200\)</EquationSource> </InlineEquation>, with inference times under 2 seconds. Meanwhile here, the usage of VRAM is reduced by up to <InlineEquation ID="IEq12"> <EquationSource Format="TEX">\(22.4\times\)</EquationSource> </InlineEquation>, and the model disk size is under 1 MB. Experiments on devices, including a high performance computing (HPC) node, a personal computer (PC), and three smartphones, demonstrate that the optimal strategy depends on the scale of the problem: local processing for <InlineEquation ID="IEq13"> <EquationSource Format="TEX">\(N=12\)</EquationSource> </InlineEquation>, remote training with local inference for <InlineEquation ID="IEq14"> <EquationSource Format="TEX">\(N=100\)</EquationSource> </InlineEquation>, and full remote processing for <InlineEquation ID="IEq15"> <EquationSource Format="TEX">\(N=1000\)</EquationSource> </InlineEquation>.

A lightweight LSTM-based open-set RF fingerprinting identification for edge deployment

期刊论文
出版模式：

开放获取
发表日期：

2025年11月24日

查看全文

Yuxuan Hu,
Yutong Fu,
Ye Chen

摘要

With the rapid advancement of deep learning techniques, numerous neural networks have been successfully developed for radio frequency (RF) fingerprinting identification. In this work, we propose a lightweight yet reliable neural network framework featuring a 9-layer architecture based on the long short-term memory (LSTM) strategy, designed for efficient open-set fingerprinting identification. The simulated beacon frames model real-world propagation effects by incorporating random modulation, power amplifier nonlinearity, multi-path fading, inherent radio noise, and additive channel noise. We extensively evaluate the identification accuracy and efficiency of our LSTM network identification against well-known deep learning models such as ResNet (144 layers) and GoogleNet (177 layers). The evaluation covers a wide range of parameters, including transmitter variability (s), number of transmitters (N), frames per transmitter ( \(FPT\) ) and signal-to-noise ratio ( \(SNR\) ). Our results show that the LSTM network maintains an accuracy of more than \(96\%\) in \(SNR \ge {20}\) with \(FPT =200\) , even with up to \(N=1000\) transmitters. At lower values ( \(FPT \le {100}\) , \(6\le SNR \le {15}\) dB), our LSTM network outperforms GoogleNet and matches ResNet in accuracy. Furthermore, it achieves a training acceleration of up to \(116.7\times\) for \(N=100\) and \(FPT =200\) , with inference times under 2 seconds. Meanwhile here, the usage of VRAM is reduced by up to \(22.4\times\) , and the model disk size is under 1 MB. Experiments on devices, including a high performance computing (HPC) node, a personal computer (PC), and three smartphones, demonstrate that the optimal strategy depends on the scale of the problem: local processing for \(N=12\) , remote training with local inference for \(N=100\) , and full remote processing for \(N=1000\) .