Few-Shot Automatic Modulation Classification via Semi-Supervised Metric Learning and Lightweight Conv-Transformer Model

Deep learning (DL) techniques have exhibited considerable advantages in automatic modulation classification (AMC) tasks, achieving high classification accuracy. However, most DL-based AMC methods require numerous labeled samples as training dataset, which is usually hard to obtain and maybe impossible in practical non-cooperative communications. To address this issue, a novel semi-supervised metric learning and lightweight convolutional transformer (conv-transformer) network (SMCTNet) is proposed in this paper. The proposed SMCTNet firstly adopts sample enhancement techniques to improve the diversity and number of training samples by randomly selecting a large number of subtasks to support model optimization with a small number of samples. Secondly, SMCTNet reduces the difficulty of model optimization by utilizing a parameter-free metric that directly calculates the similarity between output characteristics for classification. Further, a semi-supervised training method is proposed for SMCTNet, which generates high-confidence pseudo-labels to guide the model optimization through unlabeled data to improve the generalization and recognition accuracy. In addition, SMCTNet utilizes a lightweight conv-transformer network to extract local and global information of augmented samples, and its lightweight design effectively prevents the overfitting phenomenon when there are insufficient samples. Experimental results on an open-source dataset demonstrate that SMCTNet outperforms contrastive methods by 4.45%–37.14% in recognition accuracy when only 10 labeled samples are available.