Specific Emitter Identification by Edge Pattern Detection and Incremental Open-World Learning.

Area of Science:

• Computer Science
• Signal Processing
• Machine Learning

Background:

• Specific emitter identification (SEI) uses wireless device signals to identify individuals.
• Deep learning (DL) models automatically learn features from time-domain signals for SEI.
• Existing models struggle with open-world scenarios where new device classes emerge over time.

Purpose of the Study:

• To propose an incremental open-world learning (IOWL) framework for continually recognizing and learning new classes in SEI.
• To enhance open-set recognition (OSR) and maintain identification capabilities in evolving environments.

Main Methods:

• Developed a novel exemplar selection and generalization mechanism for IOWL.
• Generated a pseudo unknown dataset using edge pattern detection (EPD) and adversarial shifting for improved OSR.
• Implemented a hybrid class-incremental learning method with boundary exemplar generation to preserve prior knowledge.

Main Results:

• The proposed IOWL framework effectively recognizes and learns new classes incrementally.
• Theoretical analysis confirmed the generalization error bounds and benefits of the method.
• Numerical results on real data demonstrated superior performance compared to baseline algorithms.

Conclusions:

• The IOWL framework offers a robust solution for SEI in dynamic, open-world environments.
• The novel exemplar selection and generalization mechanism significantly improve performance.
• This approach enables continuous adaptation and learning of new device classes in SEI systems.