An Adversarial Margin Driven Rectified Prototypical Learning Method for Few-Shot Incremental Automatic Modulation Recognition

doi:10.1109/LWC.2025.3641556

An Adversarial Margin Driven Rectified Prototypical Learning Method for Few-Shot Incremental Automatic Modulation Recognition

2025-12-26

0

PRE

AI

DOI：10.1109/LWC.2025.3641556

Original

Original request for help

Save

Abstract

En 中文

Automatic modulation recognition (AMR) has been proven to be a critical component in wireless communication systems. However, the growing diversity of modulation schemes in dynamic environments, coupled with the scarcity of labeled training data, poses significant challenges for few-shot incremental AMR (FSI-AMR) in dynamically changing wireless communication environment. To address this issue, FSI-AMR offers a promising solution. In this letter, we introduce an adversarial margin-guided rectified prototypical learning (AMG-RPL) framework, which effectively alleviates catastrophic forgetting of base modulation classes while improving adaptability to novel modulation types. The proposed approach integrates an adversarial margin learning mechanism with dual classifiers during base training, enabling the extraction of transferable modulation-invariant features while retaining discriminative modulation-specific characteristics. Furthermore, a rectified prototype learning strategy is designed to refine class prototypes in incremental sessions, minimizing cognitive bias and facilitating rapid few-shot adaptation. Comprehensive experimental evaluations demonstrate that AMG-RPL outperforms existing advanced methods in FSI-AMR tasks.

Keywords:

Few-shot incremental automatic modulation recognition

adversarial margin

rectified prototypical learning

An Adversarial Margin Driven Rectified Prototypical Learning Method for Few-Shot Incremental Automatic Modulation Recognition

Abstract

Journal

Organization