RF-Enabled Deep-Learning-Assisted Drone Detection and Identification: An End-to-End Approach

Area of Science:

• Electrical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Unmanned aerial vehicles (UAVs) present growing security and privacy concerns due to their increasing affordability and widespread use.
• Monitoring and managing UAVs in restricted areas is challenging due to their sheer numbers and signal interference from other devices.
• Existing UAV identification methods often rely on computationally intensive feature-extraction techniques.

Purpose of the Study:

• To propose an efficient, end-to-end deep learning model for detecting and identifying UAVs based on their radio frequency (RF) signatures.
• To overcome the limitations of computationally expensive traditional methods by utilizing multiscale feature extraction and residual blocks.
• To evaluate the model's performance in the presence of interfering Bluetooth and Wi-Fi signals and across various signal-to-noise ratios (SNR).

Main Methods:

• An end-to-end deep learning architecture was developed for UAV detection and identification.
• Multiscale feature-extraction techniques were employed without manual intervention to capture enriched signal features.
• Residual blocks were integrated to enhance learning of complex representations and mitigate vanishing gradient issues.
• The model was trained and tested on the CardRF dataset, considering specific device and manufacturer signatures.

Main Results:

• The proposed model achieved high performance metrics: 97.53% accuracy, 98.06% precision, 98.00% sensitivity, and 98.00% F1-score for RF signal detection across 0-30 dB SNR.
• The model demonstrated a rapid inference time of 0.37 milliseconds for RF signal detection.
• Performance was evaluated in the presence of co-existing Bluetooth and Wi-Fi signals, showcasing robustness.

Conclusions:

• The developed deep learning model offers a computationally efficient and highly accurate solution for real-time UAV detection and identification.
• The end-to-end approach, utilizing multiscale features and residual blocks, surpasses existing methods in both performance and time complexity.
• The model is suitable for integration into surveillance systems for effective monitoring of airspace.