SignalFormer: Hybrid Transformer for Automatic Drone Identification Based on Drone RF Signals

Area of Science:

• Electrical Engineering
• Computer Science
• Signal Processing

Background:

• The increasing use of drones in civilian applications necessitates robust automatic drone identification (ADI) systems.
• Existing convolutional neural network (CNN) methods for ADI face limitations due to the local connectivity of convolution operators, hindering radio frequency (RF) signal identification.
• Malicious drone activities pose significant security risks, demanding advanced detection and identification technologies.

Purpose of the Study:

• To develop an innovative hybrid transformer model for enhanced automatic drone identification (ADI).
• To address the limitations of CNNs in capturing global time/frequency correlations in RF signals for drone detection.
• To improve the accuracy and reliability of drone signal identification under challenging environmental conditions.

Main Methods:

• Proposed a hybrid transformer model, SignalFormer, integrating a CNN-based tokenization method for generating time-frequency (T-F) tokens with local context.
• Employed an efficient gated self-attention mechanism within the transformer to capture global correlations among T-F tokens.
• Incorporated phase information into the input data to enhance model performance.
• Evaluated the model on two public datasets under conditions of Gaussian white noise and co-frequency signal interference.

Main Results:

• SignalFormer achieved high identification accuracy: 97.57% and 98.03% for coarse-grained tasks, and 97.48% and 98.16% for fine-grained tasks.
• Demonstrated competence in handling previously unseen drone signal categories through class-incremental learning evaluation.
• The model showed resilience against noise and signal interference, crucial for real-world ADI applications.

Conclusions:

• The proposed SignalFormer model offers a significant advancement in automatic drone identification technology.
• The hybrid approach effectively combines local feature extraction (CNN) with global context modeling (transformer) for superior RF signal analysis.
• SignalFormer presents a promising and reliable solution for mitigating threats posed by malicious drone flights.