HET: An Efficient High-Frequency Enhancement Transformer for Specific Emitter Identification.

Area of Science:

• Signal Processing
• Machine Learning
• Wireless Communications

Background:

• Specific emitter identification (SEI) is vital for wireless communications and radar.
• Standard Transformers struggle with high-frequency fingerprint features crucial for SEI.
• Existing methods face computational challenges with high-frequency sensing.

Purpose of the Study:

• To propose a novel Transformer architecture, HET, for improved SEI performance.
• To address the limitations of low-frequency bias in Transformers for SEI.
• To reduce computational complexity while enhancing high-frequency feature extraction.

Main Methods:

• Introduced a high-frequency enhanced and low-complexity Transformer (HET).
• Integrated multihead low-complexity self-attention (MLSA) for reduced complexity.
• Incorporated high-frequency enhanced connection and multihead high-frequency enhanced low-complexity self-attention (MESA) for feature reconstruction.
• Developed three HET variants (HETn, HETu, HETm) using different MESA modules.

Main Results:

• HET variants demonstrated competitive accuracy on XSRP, ADS-B, and Wi-Fi datasets.
• The proposed models achieved faster throughput compared to existing methods.
• Theoretical analysis and frequency response confirmed HET's enhanced high-frequency gain for SEI.

Conclusions:

• HET effectively overcomes Transformer limitations in SEI by focusing on high-frequency information.
• The proposed architecture offers a balance of accuracy and computational efficiency.
• HET presents a promising advancement for specific emitter identification applications.