Robustness of Deep-Learning-Based RF UAV Detectors.

Area of Science:

• Electrical Engineering
• Computer Science
• Defense Technology

Background:

• The increasing prevalence of small, low-cost unmanned aerial vehicles (UAVs) presents a significant security challenge.
• Passive radio frequency (RF) spectrum scanning, enhanced by machine learning (ML) and deep learning (DL), is a key method for detecting UAV control signals.
• Existing research focuses on optimizing ML/DL model accuracy, often overlooking robustness across diverse operational scenarios.

Purpose of the Study:

• To address the critical gap in evaluating the robustness of UAV detection systems.
• To introduce a novel dataset specifically designed to assess detector performance under varying channel conditions.
• To compare the robustness of different ML/DL architectures, specifically Convolutional Neural Networks (CNNs) and coefficient classifiers, in challenging RF environments.

Main Methods:

• Development of a new dataset featuring multiple test data categories based on diverse channel conditions, distinct from the training data pool.
• Evaluation of existing UAV detection models, including CNNs and coefficient classifiers, using the newly created dataset.
• Quantitative analysis of classifier performance, particularly under low signal-to-noise ratio (SNR) conditions and varied RF channel characteristics.

Main Results:

• Image classifiers, specifically CNNs, demonstrated approximately 40% greater robustness compared to coefficient classifiers under low SNR conditions.
• CNN classifiers maintained consistent accuracy across various RF channel conditions not encountered during training.
• Coefficient classifiers exhibited significant performance degradation, ranging from partial to complete failure, depending on specific channel conditions.

Conclusions:

• Robustness is a critical, yet often overlooked, performance metric for UAV detection systems.
• CNN-based approaches offer superior robustness for detecting UAVs via RF signal analysis, especially in adverse channel conditions.
• The developed dataset provides a valuable resource for future research aimed at enhancing the reliability of counter-UAV technologies.