Automatic modulation recognition (AMR) is essential for ensuring the physical-layer security for Internet of things (IoT) networks. Despite advancements in deep learning, most current AMR methods rely heavily on a large number of labeled samples to achieve high recognition accuracy. However, acquiring labeled samples can be costly and impractical in many real-world scenarios due to privacy concerns and economic constraints. In contrast, unlabeled data is often abundant and readily available. This paper presents a novel semi-supervised AMR framework that addresses the challenge of label scarcity by leveraging semantic mask contrastive learning (SMCL). Through a self-supervised modulation semantic mask contrastive prediction task within IQ sequence, our method learns subtle modulation features directly from unlabeled radio signals. It is important to note that SMCL requires neither data augmentation nor representation domain transformation. Sufficient experiments on public datasets have demonstrated our method outperforms existing semi-supervised and supervised methods when using the same number of labeled samples. SMCL effectively enables the representation learning of unlabeled radio signals, overcoming the limitations posed by the lack of sufficient labeled data and providing a solid technical foundation for the development of signal-based IoT large language models (IoT-LLMs).

Automatic modulation recognition (AMR) is critical in modern wireless communication systems and cognitive radio applications. Deep learning (DL) methods have become main-stream for AMR, achieving remarkable performance on balanced datasets. However, practical scenarios commonly exhibit long-tailed distributions, where abundant samples exist for head modulation types but middle and tail types suffer from severe scarcity. Such imbalance biases standard DL models towards the head classes, significantly compromising their effectiveness on tail classes. Current signal balance strategies are limited by domain-specific knowledge and insufficient sample diversity. To address these challenges, this paper proposes DiffuMLR, a novel class-dependent label information attenuation diffusion model for long-tailed modulation recognition. Specifically, we propose a label decay mechanism within a diffusion probabilistic model, dynamically adjusting label contributions during the reverse denoising process. DiffuMLR ensures semantic invariance of tail class samples while enhancing sample diversity. Extensive experiments conducted on real-world and public datasets with varying imbalance factors demonstrate significant improvements in recognition accuracy under long-tailed conditions. DiffuMLR can seamlessly integrate with existing AMR architectures and enhance AMR model robustness in realistic wireless communication scenarios.

Existing deep learning (DL)-based AMR methods achieve impressive performance with abundant labeled signals. However, in non-cooperative scenarios, acquiring labeled signals is challenging, leaving a vast amount of unlabeled data underutilized by current DL based automatic modulation recognition (AMR) approaches. To address this challenge, we propose a semi-supervised AMR framework based on a modulation signal diffusion generative model (MSDGM). This framework follows a two-stage training strategy. In the first stage, unlike existing unsupervised paradigms that rely on proxy tasks or pseudo-labels, MSDGM learns the data distribution from unlabeled signals. In the second stage, MSDGM is frozen as a feature extractor, and a classifier is trained on a small set of labeled signals to achieve effective AMR. This decoupled training paradigm significantly reduces the dependence on label quantity, ensuring rapid generalization and robust recognition even with minimal labeled signals. Extensive experimental results demonstrate the superior performance of the proposed MSDGM-based framework under few labeled signal scenarios. Notably, at SNR>0dB with only 2 labeled signals for each class, the proposed method achieves accuracies exceeding 72% and 76% on the 11-modulation-type recognition tasks of RML2016.10A and RML2022, respectively, significantly outperforming existing supervised and semi-supervised AMR methods.

Automatic Modulation Recognition (AMR) plays a pivotal role in modern mobile communications and the advancement of B5G and 6G technologies. However, the progressively intricate and hostile electromagnetic environments pose challenges to modulation recognition. While deep learning can solve complex problems, Digital Signal Processing (DSP) is interpretable and can be more computationally efficient. To combine both, we propose a Wavelet-based Adaptive modulation recognition Network (WAN) specifically designed for low SNR conditions. Diverging from traditional methods that preprocess signals prior to neural network input, our novel approach facilitates mutual synergy between DSP and the neural network during the training phase. We introduce two sub-blocks [Wavelet Threshold Estimate Block (WTEB), Selective Multi-scale Feature Extraction Block (SMFB)], which enable adaptive wavelet transform utilization for extracting multi-scale modulation features from recovery signals. WAN significantly enhances modulation recognition accuracy in low SNR while concurrently augmenting the interpretability of the neural network. Experimental results demonstrate that WAN outperforms SOTA methods in recognition accuracy.

Automatic modulation classification (AMC), as a key technology of cognitive radio, has become a focal point of research. However, most deep learning-based AMC methods require an extensive number of labeled signals to acquire a comprehensive understanding of modulation types, placing substantial pressure on signal acquisition and labeling. To solve this issue, we propose a few-shot AMC (FSAMC) method to facilitate rapid generalization and recognition with limited data, namely multi-scale feature fusion and distribution similarity network (MS2F-DS). Firstly, we design a multi-scale feature fusion (MS2F) model, which aims to extract features with varying fields of view and boost feature fusion, enabling the derivation of contextual information from the signal. Furthermore, we introduce a distribution similarity (DS) classifier to address the insufficient measurement of current similarity measurement functions by considering both micro and macro perspectives of vectors, further increasing intra-class compactness and inter-class separability. Finally, extensive experiments were conducted on 3-way 1, 3, and 5-shot FSAMC tasks using public datasets RML2016.10a and RML2016.10b, and the results demonstrated the effectiveness of our method.

Recently, significant progress has been made in deep learning, which has been widely applied in automatic modulation classification (AMC) with remarkable outcomes. However, current deep learning based AMC (DL-AMC) algorithms show limitations in their ability to accommodate dynamically changing communication scenarios. With the increasing number of modulation types, most DL-AMC algorithms often need to be re-trained, making it hard to transfer previous knowledge to the new models. Also, modulation classification faces the difficulty of acquiring and annotating a large number of signals. To address these challenges, we have modeled a few-shot class-incremental AMC (FSCI-AMC) task and proposed a pseudo classes and stochastic classifiers-based network (PASS-Net) to accomplish it. Firstly, the pseudo classes are generated to reserve space for new types, enhancing the model’s continuous learning capability. Additionally, stochastic classifiers ensure the reliability of generated pseudo classes. Finally, in the incremental session, both real and pseudo classes are used for modulation classification. To evaluate the proposed approach, experiments were conducted on 7 modulation types as base classes and another 7 modulation types as incremental classes. The results show superior performance in 7 sessions of 1-way 1-shot and 1-way 5-shot class-incremental experiments compared to other competitive methods.

It is difficult to realize effective synthetic aperture radar (SAR) automatic target recognition (ATR) in open scenarios because the ATR model cannot continuously learn from new classes with limited training samples. When adding new classes to the previously trained model, the capability of recognizing old classes may lose due to severe overfitting. To tackle this problem, a few-shot class-incremental SAR ATR method, namely, hierarchical embedding and incremental evolutionary network (HEIEN), is proposed in this article. First, a hierarchical embedding network and a hybrid distance-based classifier are constructed for basic feature extraction and classification. Then, in order to obtain more accurate decision boundaries, an adaptive class-incremental learning (ACIL) module is designed to adjust the weights of classifiers in all tasks by collecting context information from the past to the present. Finally, a pseudo-incremental training strategy is designed to enable effective model training with only a few samples. Experimental results on the moving and stationary target acquisition and recognition (MSTAR) benchmark data set have illustrated that HEIEN performs well with remarkable advantages in few-shot class-incremental SAR ATR tasks.