A Self-Distillation Framework for Few-Shot Automatic Modulation Recognition

Recently, various deep learning (DL) methods have effectively improved the accuracy of automatic modulation recognition (AMR). However, in practical applications, few labeled samples can be obtained. Therefore, recognizing signals with a small number of labeled samples (few-shot learning) is both crucial and challenging. In this paper, we propose a self-distillation framework based on dual domain feature fusion (SD-DuDo). In this framework, we propose a dual-domain encoder capable of fusing sequence-domain and frequency-domain features. Self-distillation pre-training of the dual-domain encoder is performed using unlabeled signal. Then, the pre-trained dual domain encoder and a classifier are fine-tuned using few labeled samples. The proposed framework is compared with existing few-shot methods using a small amount of labeled data from the RML2016.10a dataset to demonstrate its superiority and stability.