PaperlayerView exact revision on arXiv

Source-linked AI summary

Radio Frequency Fingerprint Identification for LoRa Using Spectrogram and CNN

Guanxiong Shen, Junqing Zhang, Alan Marshall, Linning Peng, Xianbin Wang

arXiv:2101.01668v1eess.SPcs.LG

TL;DR

The paper tackles LoRa device authentication with RFFI, focusing on the limitations of signal representation and CFO-related instability. It uses spectrograms with CNN classification, compensates CFO drift, and calibrates CNN outputs with a hybrid classifier. In experiments with 20 LoRa devices, the spectrogram-based scheme reached 97.61% accuracy.

  • Problem

    LoRa RFFI needs a representation suited to chirp modulation and must address CFO drift that can cause misclassification and compromise stability.

  • Method

    The paper uses spectrogram representations with CNN classification, CFO compensation, and a hybrid classifier that calibrates CNN outputs using estimated CFO.

  • Results

    97.61% accuracy was achieved for 20 LoRa devices with the spectrogram-based hybrid scheme, exceeding the CNN-only spectrogram result of 96.44%.

  • Takeaways & Limitations

    Spectrograms expose LoRa time-frequency characteristics, while CFO compensation and CFO-based calibration improve RFFI performance and stability.

Abstract

from arXiv · show

Radio frequency fingerprint identification (RFFI) is an emerging device authentication technique that relies on intrinsic hardware characteristics of wireless devices. We designed an RFFI scheme for Long Range (LoRa) systems based on spectrogram and convolutional neural network (CNN). Specifically, we used spectrogram to represent the fine-grained time-frequency characteristics of LoRa signals. In addition, we revealed that the instantaneous carrier frequency offset (CFO) is drifting, which will result in misclassification and significantly compromise the system stability; we demonstrated CFO compensation is an effective mitigation. Finally, we designed a hybrid classifier that can adjust CNN outputs with the estimated CFO. The mean value of CFO remains relatively stable, hence it can be used to rule out CNN predictions whose estimated CFO falls out of the range. We performed experiments in real wireless environments using 20 LoRa devices under test (DUTs) and a Universal Software Radio Peripheral (USRP) N210 receiver. By comparing with the IQ-based and FFT-based RFFI schemes, our spectrogram-based scheme can reach the best classification accuracy, i.e., 97.61% for 20 LoRa DUTs.

I. INTRODUCTION

The paper addresses LoRa device authentication by combining RFFI with LoRa-specific spectrogram representations and CNNs, while investigating CFO-driven instability. Experiments compare signal representations and introduce CFO compensation and hybrid classification to improve performance.

  • I. INTRODUCTION: LoRa RFFI remains challenging because prior work did not fully exploit LoRa’s time-frequency modulation or investigate CFO variation.LoRa uses chirp spread spectrum, while low-cost oscillators can introduce CFO changes.
  • I. INTRODUCTION: Spectrogram reached 96.44% accuracy, outperforming IQ samples at 83.36% and FFT results at 87.36%.Spectrogram-based training also took about 20 minutes versus one hour for IQ/FFT models.
  • I. INTRODUCTION: CFO varies over short periods, degrades RFFI stability, and compensation improved spectrogram-CNN accuracy from 75.59% to 96.44%.The authors used an attenuated LoRa-DUT-to-USRP setup to reduce channel effects while examining CFO variation.
  • I. INTRODUCTION: The hybrid classifier calibrates CNN softmax outputs using estimated CFO, improving IQ-based accuracy from 83.36% to 92.01% in the best case.The calibration uses CFO stability over longer terms to rule out inconsistent CNN predictions.

A. LoRa Modulation Technique

LoRa uses linear chirps whose time-frequency structure can be represented with spectrograms. The paper uses STFT-based processing to expose changing instantaneous frequency and signal parameters from LoRa preambles.

  • A. LoRa Modulation Technique: LoRa CSS encodes symbols with linear chirps whose frequency increases or decreases linearly over time.The information in each symbol is encoded in the chirp’s initial phase.
  • A. LoRa Modulation Technique: The preamble is identical in every packet and for every LoRa device type, making it suitable for consistent signal representation.The paper illustrates one preamble in both time-domain and spectrogram forms.
  • B. Short-Time Fourier Transform and Spectrogram: STFT divides a signal into short segments and applies a Fourier transform to each segment.Its inputs include the analyzed signal and window function; the hop size controls segment displacement.
  • B. Short-Time Fourier Transform and Spectrogram: The spectrogram is the squared magnitude of the STFT, representing how instantaneous frequency changes over time.It also exposes parameters such as bandwidth and symbol duration.

III. LORA RECEIVER OPERATION

The LoRa receiver converts the transmitted RF signal into a sampled digital baseband signal. Carrier frequency mismatch between transmitter and receiver appears as CFO in that received signal.

  • III. LORA RECEIVER OPERATION: The receiver antenna captures the RF signal, which is mixed down to baseband and sampled by an ADC.The resulting digital baseband signal is represented at sampling intervals.
  • III. LORA RECEIVER OPERATION: CFO is the frequency difference between the transmitter and receiver carrier frequencies.This offset appears in the received baseband signal and is central to the paper’s later compensation procedure.
  • III. LORA RECEIVER OPERATION: The receiver-operation section introduces CFO estimation and compensation for LoRa signals.The compensated signal is intended to account for the offset affecting instantaneous frequency.

1) Coarse CFO Estimation:

The received LoRa chirp has an inevitable CFO that shifts its instantaneous frequency, so the system estimates and compensates this offset in coarse and fine stages.

  • 1) Coarse CFO Estimation:: The ideal chirp frequency increases linearly, but CFO adds an offset to the received instantaneous frequency.This linearity enables CFO estimation from received preambles.
  • 1) Coarse CFO Estimation:: The coarse CFO estimate is obtained from the mean instantaneous frequency of the received preambles, using the symbol length L.The received signal is then compensated with the estimated frequency offset.
  • 1) Coarse CFO Estimation:: Residual CFO remains after coarse compensation, so a fine estimator exploits the repeating structure of LoRa preambles before further compensation.The fine estimate uses the phase relationship between samples separated by one symbol length.
  • 1) Coarse CFO Estimation:: The fine estimator is limited by phase ambiguity, with |∆bffine| < B/2^(SF+1).For SF=7 and B=125 kHz, the estimation range is ±488.3 Hz, whereas an 868 MHz oscillator drift of ±10 ppm is approximately 8.68 kHz.

IV. RFFI SYSTEM

The RFFI system synchronizes packets, estimates and compensates CFO, and records CFO values for later hybrid classification. CFO compensation is included because its effect on low-cost IoT RFFI had not been experimentally investigated.

  • IV. RFFI SYSTEM: Synchronization locates LoRa packets from repeated preambles before subsequent RFFI processing.This is described as a standard communication-system procedure.
  • IV. RFFI SYSTEM: The system performs CFO estimation and compensation, addressing prior RFFI studies that either used raw IQ samples or treated CFO as a feature.The effect of CFO on low-cost IoT RFFI had not been experimentally investigated.
  • IV. RFFI SYSTEM: During training, the system builds a CFO database containing estimated CFO values for each device, which supports the later hybrid classifier.The database is generated from the adopted CFO estimation and compensation algorithms.

B. Normalization

The system normalizes received signals, uses preambles for representation learning, and compares IQ, FFT, and spectrogram inputs before CNN classification.

  • B. Normalization: Signal normalization removes power differences because received power depends on distance rather than device identity.The normalized signal is defined using the root mean square amplitude.
  • B. Normalization: Only the preamble is used so the model does not learn protocol-specific or data-specific knowledge.Signal representation is intended to expose underlying characteristics that the classifier can learn.
  • B. Normalization: IQ represents time-domain signals, whereas FFT converts them into frequency-domain features that may be less obvious in time-domain data.These representations provide alternative inputs for RFFI classification.
  • B. Normalization: Spectrograms represent time-frequency characteristics by showing both frequency information and its evolution over time.The paper applies logarithmic magnitude compression as spectrogram preprocessing.
  • B. Normalization: CNN convolutional and pooling layers extract features automatically, while fully connected layers classify the extracted representations.The final CNN softmax layer outputs class probabilities, and the highest-probability class is normally selected.

E. Hybrid Classifier

The hybrid classifier uses CFO consistency to filter unreliable CNN predictions because similar devices can produce close class probabilities. It then selects the highest remaining probability.

  • E. Hybrid Classifier: CNN outputs can misclassify similar devices when their class probabilities are close, such as S1=0.51 and S2=0.49.Selecting only the largest CNN probability is unreliable in this situation.
  • E. Hybrid Classifier: The hybrid classifier compares each DUT’s estimated CFO with a reference CFO database built for all K devices.The method is motivated by relatively stable mean CFO values across different days.
  • E. Hybrid Classifier: When the CFO difference exceeds the predefined threshold, the classifier sets that device’s CNN probability to zero; otherwise, it preserves the probability.The final label is the device with the highest probability after calibration.
  • E. Hybrid Classifier: The spectrogram-based CNN is the model architecture used for this hybrid classification, while the paper also presents a separate IQ/FFT-based CNN architecture.The spectrogram model uses convolutional layers followed by fully connected classification and softmax probabilities.

B. IQ/FFT-based CNN

The study compares CNN inputs based on IQ samples, FFT results, and spectrograms under matched network settings, while examining the experimental setup used to study CFO effects.

  • B. IQ/FFT-based CNN: The IQ/FFT-based CNN uses complex signals as two independent dimensions, allowing the same CNN architecture to process IQ and FFT data.Its architecture has three convolutional layers, batch normalization, ReLU activations, and max-pooling layers.
  • B. IQ/FFT-based CNN: 20 minutes versus one hour: the spectrogram-based model trains faster than the IQ/FFT-based model under the same settings.The spectrogram model's loss also drops earlier and faster.
  • B. IQ/FFT-based CNN: The CFO study used ten LoRa devices, a USRP N210 receiver, and datasets collected over four days to isolate and evaluate CFO variation.The devices transmitted for about one hour per day, producing 3,000 packets per device per day.

B. CFO Drift

CFO varies both shortly after device startup and across collection days, destabilizing CNN-only RFFI across mismatched days; compensation restores accuracy above 96% across four days.

  • B. CFO Drift: CFO decreases during the first 20 minutes after startup and then remains relatively constant within the one-hour observation period.The paper relates this short-time pattern to self-heating and oscillator sensitivity to temperature.
  • B. CFO Drift: Average CFO changes non-negligibly and unpredictably across days but remains relatively stable enough to rule out predictions outside a device's CFO range.The Day 1-to-Day 4 comparison links hundreds-of-hertz CFO shifts for Dev 3 and Dev 5 with confusion toward other devices.
  • B. CFO Drift: 99.57% on same-day testing fell to 78.84%, 85.32%, and 77.83% on Days 2–4 without CFO compensation.Devices such as Dev 3 and Dev 5 were completely misclassified in the worst cross-day case.
  • B. CFO Drift: Above 96% accuracy across four days: CFO compensation prevented the degradation observed in the uncompensated CNN-only classifier.The compensated accuracies were 98.89%, 98.05%, 96.73%, and 96.93% from Day 1 through Day 4.

VII. EXPERIMENTAL EVALUATIONS IN A REAL WIRELESS ENVIRONMENT

In a real indoor wireless environment, the study evaluates signal representations, CFO compensation, and hybrid calibration using 20 LoRa devices. The spectrogram-based CNN provides the strongest representation result among the reported CNN-only comparisons.

  • VII. EXPERIMENTAL EVALUATIONS IN A REAL WIRELESS ENVIRONMENT: The evaluation examined representation choice, CFO effects in wireless conditions, and calibration by the proposed hybrid classifier.The experiments used separate training and test packets from each device.
  • VII. EXPERIMENTAL EVALUATIONS IN A REAL WIRELESS ENVIRONMENT: 96.44% versus 83.36% and 87.36%: the spectrogram-based CNN outperformed the IQ- and FFT-based CNNs in overall accuracy.The comparison used similar network structures for the three signal representations.
  • VII. EXPERIMENTAL EVALUATIONS IN A REAL WIRELESS ENVIRONMENT: About 20 minutes versus one hour: the spectrogram-based model required less training time than the IQ/FFT-based model.The spectrogram model's loss decreased earlier and faster.

C. Impact of CFO drift

CFO drift changes the distribution between training and test packets, degrading classification unless compensated. CFO-aware calibration then further improves predictions because average CFO values remain relatively stable.

  • C. Impact of CFO drift: 96.44% accuracy followed CFO compensation for spectrogram-based RFFI, up from 75.59% without compensation.IQ-based accuracy rose from 59.44% to 83.36%, while FFT-based accuracy rose from 51.62% to 87.36%.
  • C. Impact of CFO drift: Different CFO distributions between training and test packets produced unacceptable cross-day classification despite 99.57% same-day accuracy without compensation.The test packets had CFOs different from the training packets, creating a distribution mismatch.
  • C. Impact of CFO drift: t-SNE visualizations separated training and test packets into distinct clusters without CFO compensation, whereas compensation altered this separation.The visualization represents 2,000 Dev 1 packets, with training and test packets shown separately.
  • C. Impact of CFO drift: CFO varies across packets and over time, but its average remains stable enough to rule out predictions whose estimated CFO differs substantially from the reference.The hybrid classifier calibrates CNN softmax outputs using the estimated CFO rather than treating CFO alone as a device fingerprint.
  • C. Impact of CFO drift: 97.61% spectrogram accuracy was achieved after hybrid CFO calibration, compared with 96.44% for the CNN-only classifier after compensation.The hybrid classifier improved all three signal representations, including FFT from 87.36% to 92.31%.
  • C. Impact of CFO drift: The hybrid classifier provides no additional benefit without CFO compensation because CFO already contributes to the uncorrected CNN prediction.This limitation makes compensation a prerequisite for the hybrid calibration strategy.
Loading 2101.01668v1…