PaperlayerView exact revision on arXiv

Source-linked AI summary

Micro-UAV Detection and Classification from RF Fingerprints Using Machine Learning Techniques

Martins Ezuma, Fatih Erden, Chethan Kumar Anjinappa, Ozgur Ozdemir, Ismail Guvenc

arXiv:1901.07703v2eess.SP

TL;DR

The paper addresses micro-UAV detection and controller identification using RF fingerprints despite limitations of conventional and time-domain approaches. It combines wavelet-domain Markov-model detection with energy-transient feature classification, achieving 96.3% average kNN accuracy across 14 controllers.

  • Problem

    Existing micro-UAV detection and classification methods face low radar cross-sections, limited vision and acoustic range, environmental noise, and unreliable time-domain transient detection.

  • Method

    The method uses wavelet-domain Markov-model naive Bayes for detection and NCA-selected statistical features from energy transients for machine-learning classification.

  • Results

    96.3% average classification accuracy is achieved with kNN across 14 micro-UAV controllers, while all micro-UAVs are detected correctly.

  • Takeaways & Limitations

    RF fingerprints can support micro-UAV detection and classification, with kNN and SVM maintaining accuracy above 95% as the number of controllers increases within a certain level.

Abstract

from arXiv · show

This paper focuses on the detection and classification of micro-unmanned aerial vehicles (UAVs) using radio frequency (RF) fingerprints of the signals transmitted from the controller to the micro-UAV. In the detection phase, raw signals are split into frames and transformed into the wavelet domain. A Markov models-based naive Bayes approach is used to check for the presence of a UAV in each frame. In the classification phase, unlike the traditional approaches that rely solely on time-domain signals and corresponding features, the proposed technique uses the energy transient signal. This approach is more robust to noise and can cope with different modulation techniques. First, the normalized energy trajectory is generated from the energy-time-frequency distribution of the raw control signal. Next, the start and end points of the energy transient are detected by searching for the most abrupt changes in the mean of the energy trajectory. Then, a set of statistical features is extracted from the energy transient. Significant features are selected by performing neighborhood component analysis (NCA) to keep the computational cost of the algorithm low. Finally, selected features are fed to several machine learning algorithms for classification. The algorithms are evaluated experimentally using a database containing 100 RF signals from each of 14 different UAV controllers. The signals are recorded wirelessly using a high-frequency oscilloscope. The data set is randomly partitioned into training and test sets for validation with the ratio 4:1. Ten Monte Carlo simulations are run and results are averaged to assess the performance of the methods. All the micro-UAVs are detected correctly and an average accuracy of 96.3% is achieved using the k-nearest neighbor (kNN) classification. Proposed methods are also tested for different signal-to-noise ratio (SNR) levels and results are reported.

I. INTRODUCTION

Micro-UAV proliferation and associated public-safety threats motivate reliable detection and classification, but conventional radar, vision, sound, and time-domain RF methods face important limitations. The paper proposes RF-based detection using wavelet-domain Markov-model naive Bayes and classification using energy-transient features selected by NCA.

  • Micro-UAVs support applications including precision agriculture, infrastructure monitoring, package delivery, entertainment, and ad hoc Internet connectivity.
  • Public-safety risks include privacy violations, sensitive-facility intrusions, safety incidents, and attacks involving explosives or CBRNE materials.
  • Conventional radar, sound, video, and time-domain RF techniques face detection limitations caused by low radar cross-section, short range, environmental noise, or gradual signal transitions.
  • The proposed classifier transforms signals into the energy-time-frequency domain, extracts statistical energy-transient features, reduces dimensionality with NCA, and applies machine-learning classifiers.
  • The proposed detector transforms RF signals into the wavelet domain and uses Markov-model-based naive Bayes to distinguish noise from micro-UAV signals.

B. Vision-based Techniques

Vision-based micro-UAV detection uses engineered image features or learned CNN representations, while related acoustic systems use sound signatures and multimodal sensing. These approaches face practical constraints from environmental conditions, data requirements, sensing range, and surveyed-area size.

  • Vision-based detection: High-resolution cameras support micro-UAV detection by extracting Haar-like, HOG, and LBP features for cascaded boosted classifiers.The classifiers operate in multi-stage sequences with increasing complexity.
  • Vision-based detection: CNN-based methods learn detection features autonomously rather than relying on human-crafted features, but require large amounts of training data.
  • Vision-based detection: GFD-based computer vision uses SURF keypoints and shape descriptors, providing faster micro-UAV detection than CNN methods.
  • Limitations: Vision-based detection depends on ambient conditions such as lighting and may perform poorly when the surveyed area is large.
  • Acoustic and hybrid techniques: Acoustic and hybrid systems identify micro-UAVs using motor-related sound signatures, audio features, or combined camera and microphone data with classifiers such as SVMs.Reported systems use microphone arrays, MFCCs, HOG features, SVMs, GMMs, and deep learning algorithms for detection, identification, localization, or tracking.
  • Limitations: Audio-based systems are constrained by the practical range of commercial microphones.

D. RF fingerprinting

RF fingerprinting uses passive reception of controller signals to detect and identify micro-UAVs, addressing limitations of radar, vision, and acoustic methods. The paper combines state-transition-based detection with energy-trajectory features and machine-learning classification.

  • RF fingerprints exploit controller-specific circuitry and modulation characteristics to detect and classify micro-UAVs.
  • Passive RF reception avoids transmission, while sensitive receivers, high-gain antennas, and denoising can address low RCS, range, and environmental-noise challenges.
  • The system detects controller transmissions and then identifies the UAV type from extracted RF fingerprints.
  • Detection uses state-transition probabilities rather than a single fixed energy threshold to reduce false alarms from random background-noise bursts.
  • Classification extracts energy-trajectory features from spectrograms, selects significant features with NCA, and applies machine-learning models.

IV. UAV DETECTION USING RF SIGNALS

The detection stage captures controller waveforms and distinguishes UAV transmissions from noise and other same-band wireless signals. Wavelet preprocessing is used to improve detection under low-SNR conditions.

  • RF signals from different micro-UAV controllers exhibit distinct waveforms linked to transmitter circuits, modulation, and packet structure.
  • Simple thresholding is unreliable in noisy environments, and time-domain transient extraction is difficult for fingerprint generation.
  • The detector classifies received signals as UAV or non-UAV, with non-UAV signals represented by receiver noise and same-band interference.
  • Wavelet-domain preprocessing is applied before detection to help identify RF signals in low-SNR regimes.

A. Pre-processing Step

Wavelet preprocessing improves received-signal SNR through denoising while compressing data without information loss, supporting faster and lower-complexity detection.

  • Wavelet transforms provide denoising and data compression without information loss compared with traditional time- and Fourier-domain analysis.
  • Improved SNR supports detection capability, while compression supports low-complexity design and faster detection algorithms.
  • The implementation uses three-stage Haar wavelet decomposition with low-pass and high-pass filters followed by downsampling.

B. Bayesian Decision Making

The detection stage models transformed RF signals as state sequences and uses class-specific transition probabilities within a Bayesian decision framework to distinguish UAV signals from noise.

  • Preprocessing: Wavelet coefficients after the third decomposition level provide the detection algorithm’s training and testing inputs.The decomposition uses simple Haar low-pass and high-pass filters followed by downsampling.
  • State representation: Three states are assigned to each transformed-signal sample using thresholds T1 and T2, producing a state sequence for detection.The thresholds separate samples above T1, between T2 and T1, and below T2.
  • Markov modeling: 3-state Markov models capture dependencies between adjacent signal states for the UAV and noise classes.Transition probabilities are calculated from adjacent indexed states and organized into class-specific matrices.
  • Markov modeling: Transition probabilities are obtained by normalizing counts of transitions from state Si to Sj by the total number of samples.The transition-count matrix TN records each Nij transition count, while TP contains the normalized probabilities pij.
  • Bayesian decision: Thresholds T1 and T2 are set to ±3σ of the wavelet-transformed environmental noise signal.This choice is motivated by modeling environmental noise as Gaussian, with approximately 0.993 of samples expected within the ±3σ band.
  • Bayesian decision: The Bayesian detector compares UAV and noise likelihoods, favoring C = 1 when log(P(yT |C = 1)) ≥ log(P(yT |C = 0)).The posterior formulation uses the transformed signal yT, class likelihoods, priors, and evidence.

V. UAV CLASSIFICATION USING RF FINGERPRINTS

The classification stage represents RF signals through energy-time-frequency transients, extracts statistical fingerprints, selects significant features with NCA, and classifies them using machine learning algorithms.

  • Energy-domain representation: Energy transients in the energy-time-frequency domain are used as fingerprints of the corresponding RF signals.The approach uses a spectrogram representation rather than relying only on time-domain transients.
  • Energy-domain representation: The spectrogram uses squared-magnitude STFT values to represent signal energy across time and frequency.The representation can also reveal transmit frequency and frequency-hopping patterns.
  • Transient extraction: The energy trajectory is computed from maximum spectrogram energy values along the time axis, and the transient is located at abrupt changes in its mean or variance.The resulting transient is represented as fE(n), n = 1, . . . , N.
  • Feature extraction: Four statistical fingerprints describe the energy transient: skewness, variance, energy spectral entropy, and kurtosis.They characterize asymmetry, spread, Shannon entropy, and transient sharpness or flatness.
  • Feature selection and classification: NCA selects significant features before training kNN, discriminant analysis, SVM, and neural-network classifiers.Feature selection reduces redundancy and the computational cost of classification.

A. Feature Selection Using NCA

NCA reduces the feature set by learning a lower-dimensional projection and ranking features by importance, helping reduce computational cost while retaining discriminative features.

  • A. Feature Selection Using NCA: NCA learns a transformation matrix that projects the original feature data into a lower-dimensional space.The method is non-parametric, embedded, and supervised.
  • A. Feature Selection Using NCA: NCA maximizes a regularized objective based on leave-one-out classification probability to learn feature weights.λ is the regularization term, N is the number of samples, and p_i is the average leave-one-out probability.
  • A. Feature Selection Using NCA: The regularization term drives weights of redundant or correlated features toward zero, enabling feature selection.This avoids assuming a parametric structure in the lower-dimensional space.
  • A. Feature Selection Using NCA: Kurtosis received the highest weight, whereas skewness received the lowest weight in the reported RF-fingerprint test case.The authors attribute this relationship to correlation between skewness and kurtosis.
  • A. Feature Selection Using NCA: Discarding skewness can preserve good classifier results while reducing overfitting risk and training cost for larger classification problems.The authors recommend retaining the fewer significant features selected by NCA.

A. Experimental Setup

The experiments use wireless RF recordings from 14 micro-UAV controllers in near-field and far-field indoor settings, with controlled training-test partitioning and multiple evaluation conditions.

  • A. Experimental Setup: RF signals were collected indoors from 14 micro-UAV controllers operating at 2.4 GHz in near-field conditions.The environment included strong interference from WiFi, Bluetooth, and a microwave oven.
  • A. Experimental Setup: The measurement system used a 6 GHz-bandwidth oscilloscope, a 2 dBi omnidirectional antenna, and a 24 dBi Wi-Fi grid antenna.The antennas operated in the 2.4 GHz frequency band.
  • A. Experimental Setup: The omnidirectional antenna captured close-range signals, while the grid antenna captured far-field signals under reduced-SNR conditions.Near-field SNR was about 30 dBi and decreased with receiver distance.
  • A. Experimental Setup: Each controller contributed 100 RF signals, with each signal represented as a 5000000 × 1 vector spanning 0.25 ms.The signals were captured by an antenna and fed into the oscilloscope for recording.
  • A. Experimental Setup: The database used a 4:1 random partition, assigning 80% of the data to training and 20% to testing.The partition parameter was p = 0.2.

B. Results

Detection becomes more accurate at higher SNR, while classification favors kNN, SVM, and DA over NN, with performance affected by noise and controller count.

  • B. Results: 84% detection accuracy was achieved at SNR=10 dB, while all UAVs were detected at SNR values beyond 12 dB.The 10 dB condition corresponded to a distance of 80 m.
  • B. Results: 96.3% classification accuracy was achieved by kNN, compared with 96.84% for SVM, 88.15% for DA, and 58.49% for NN.These values were averaged over 10 Monte Carlo simulations for 14 micro-UAV controllers.
  • B. Results: At SNR=25 dB, kNN achieved about 97.29% classification accuracy, whereas NN achieved 57.14%.At SNR=10 dB, kNN, DA, and SVM reached 60-70%, while NN remained below 50%.
  • B. Results: At SNR of 25 dB, kNN, DA, and SVM remained almost unchanged as the number of controllers increased, while NN became unstable at 8 or more controllers.The authors judged NN unsuitable unless there were 6 or fewer controllers with this feature set.
  • B. Results: The reported NN comparison is limited because its hyperparameters were not optimized, unlike those of kNN, DA, and SVM.The authors identify NN hyperparameter optimization as future work.
  • B. Results: 97.1% accuracy was obtained in a sample kNN confusion matrix using 280 test signals from 14 controllers at SNR of 25 dB.Only one or two samples from four controllers were misclassified.

VII. CONCLUSION

The paper combines Markov-model Bayesian detection with energy-time-domain RF features for micro-UAV classification and reports strong kNN performance across tested conditions.

  • VII. CONCLUSION: The detection algorithm models UAV and non-UAV classes with a Bayesian approach based on Markov models.The classification method uses features extracted from the energy-time domain of RF signals.
  • VII. CONCLUSION: kNN performed best, while NN performed worst at lower SNR levels and with more controllers.The conclusion also reports kNN accuracy above 80% up to SNR of 15 dB for 14 controllers.
  • VII. CONCLUSION: kNN and SVM maintained accuracy above 95% when the number of controllers increased up to a certain level.The conclusion describes this as occurring without compromising performance.
  • VII. CONCLUSION: Future work will test outdoor environments, multiple sensors, and UAV signals relevant to low-probability-of-intercept emitters.The paper identifies such emitters as difficult to detect using a single RF sensing platform because of low peak power.
Loading 1901.07703v2…