Semi-supervised Learning Framework for UAV Detection

TL;DR

This paper introduces a semi-supervised RF signal analysis framework for UAV detection, utilizing wavelet transforms and outlier detection to achieve high accuracy in identifying UAVs based on their wireless communication fingerprints.

Contribution

The paper presents a novel semi-supervised method using wavelet packet features and outlier detection for UAV detection via RF signals, filling a gap in unsupervised UAV detection techniques.

Findings

Achieved 96.7% accuracy at 30dB SNR

Achieved 86% accuracy at 18dB SNR

Method can extend to rogue RF device detection

Abstract

The use of supervised learning with various sensing techniques such as audio, visual imaging, thermal sensing, RADAR, and radio frequency (RF) have been widely applied in the detection of unmanned aerial vehicles (UAV) in an environment. However, little or no attention has been given to the application of unsupervised or semi-supervised algorithms for UAV detection. In this paper, we proposed a semi-supervised technique and architecture for detecting UAVs in an environment by exploiting the RF signals (i.e., fingerprints) between a UAV and its flight-controller communication under wireless inference such as Bluetooth and WiFi. By decomposing the RF signals using a two-level wavelet packet transform, we estimated the second moment statistic (i.e., variance) of the coefficients in each packet as a feature set. We developed a local outlier factor model as the UAV detection algorithm using the coefficient variances of the wavelet packets from WiFi and Bluetooth signals. When detecting the presence of RF-based UAV, we achieved an accuracy of 96.7$\%$ and 86$\%$ at a signal-to-noise ratio of 30~dB and 18~dB, respectively. The application of this approach is not limited to UAV detection as it can be extended to the detection of rogue RF devices in an environment.