Automatic modulation classification (AMC) is essential in noncooperative communication systems, since it enables the automatic recognition of signal modulation types. The recent incorporation of deep learning, particularly graph neural networks (GNNs), has significantly improved the AMC accuracy. The GNNs increase the performance by decoding the relationships between nodes and edges, which represent the topological structure of data. In AMC, signal features or time points are modeled as nodes, and their interconnections represent the interactions between these features. This modeling allows the GNNs to thoroughly analyze signals and accurately identify complex modulations. However, the existing traditional methods for mapping IQ signal sequences into graphs exhibit high computational load and excessive processing time. To solve these problems, this article proposes a lightweight model of high performance, referred to as PGNet, which combines efficient partial convolution (PConv) with graph sparse attention techniques. This combination minimizes the computational load and maximizes the strengths of the convolutional neural networks and GNNs. The results of the conducted experiment show that PGNet, respectively, achieves average accuracies of 62.8% and 64.1% on the RML2016.10a and RML2016.10b datasets, with only 16315 parameters and an inference time of only 2 ms/sample. Due to its high efficiency and compact size, the proposed PGNet provides a substantial potential for deployment in low computing resource scenarios, such as IoT devices with limited resources.

Automatic modulation classification (AMC) enables the identification of modulation schemes without prior information, facilitating efficient signal processing. Recently, deep-learning (DL)-based AMC has significantly advanced signal detection and recognition across various domains, including Internet of Things (IoT) systems and industrial cognitive communication systems. While high-performing AMC models achieve remarkable accuracy, their substantial storage and computational demands hinder deployment in resource-limited IoT communication systems. To address this challenge, we propose CPPCNet, a high-performance, lightweight complex-valued partial pointwise convolutional neural network. By leveraging complex-valued operations for automatic feature extraction, CPPCNet preserves phase information, enhancing classification performance. To alleviate the computational burden of complex-valued operations in resource-limited IoTs, we introduce complex-valued partial pointwise convolution (CPPC), which optimally balances accuracy and model complexity. Experimental results show that CPPCNet, with only 65302 parameters, achieves a state-of-the-art (SOTA) accuracy of 66.38% on RML2016.10b among all existing AMC models. On RML2016.10a, it also achieves a strong performance with an accuracy of 62.25%. Furthermore, it achieves 83.5% accuracy on the HisarMod2019.1 dataset, which includes more realistic channel impairments, outperforming existing lightweight AMC models in both accuracy and inference speed. These results highlight CPPCNet’s strong generalization ability and its ability to balance performance and efficiency, making it a promising solution for AMC applications in resource-constrained and dynamically changing environments.

Communication device recognition is a key problem of electromagnetic space perception. At present, the traditional recognition technology is difficult to adapt to the complex signal situation. Thanks to the deep learning's superior capability of processing complex and massive data, it has been a hot topic in the field of communication area. However, it needs a large amount and high-quality signal dataset, which will pay for much cost. Therefore, small sample and data augmentation technology should be given much attention. In this article, we propose a novel time series data augmentation technology for small sample recognition. First, a complex neural network is designed to recognize the communication device based on the in-phase/quadrature time series. Second, based on signal data characteristic, several simple and effective methods of time series data augmentation are analyzed, which include noise disturbance, amplitude and time-delay transformation, frequency offset, and phase shift transformation. Third, in order to get a better augmentation result, based on complex neural network model characteristic, a novel data augmentation method of virtual adversarial training is presented for the small sample device recognition. Finally, a series of experimental simulations in real ADS-B signal indicates that the proposed method is suitable for the time series analysis and recognition of communication device with a small sample.

In recent years, unmanned aerial vehicles (UAV) or drones have become ubiquitous in our daily lives, bringing great convenience to our lives and playing a pivotal role in future wireless networks and the Internet of things. One of the major problems associated with the UAV is the heterogeneous nature of such deployments; this heterogeneity poses many challenges, particularly in the areas of security and privacy. The key to solving these problems is to accurately identify and authenticate drones. In this article, a reliable UAV identify framework based on radio frequency fingerprint is proposed. First, we established a wireless signal label architecture and systematically collected, analyzed, and recorded the radio frequency signals of different UAVs in different flight modes and different distances in the telemetry link, and established UAV signal datasets. Then, the intelligent algorithm and anti-UAV system are designed by using the collected dataset, and the feasibility of the developed dataset for detecting and identifying UAVs is verified by using machine learning and deep learning. The simulation results show that under the condition of Gaussian white noise, the method based on deep learning achieves high reliability, and when the SNR is not less than 5dB, the model achieves more than 95% of the monitoring and recognition accuracy. Finally, we discussed the possible applications of the dataset in the future.

Momentum term technique is an efficient resolution to accelerate the convergence speed of adaptive blind source separation (BSS) algorithms, however, the BSS algorithm combined with a momentum term also suffers from the tradeoff between the fast convergence speed and small misadjustment error. In order to alleviate such compromise, an optimal variable momentum factor method is used to boost the separating performance of the nonlinear principal component analysis (NPCA) BSS algorithm. At first, by using the projection approximation, the cost function of the NPCA algorithm can be represented as a quadratic function of the momentum factor. Then the optimal momentum factor is obtained on the basis of the gradient decent technique, which makes the cost function descend in the fastest way during each iteration. Simulation experiment results manifest that the modified algorithm can improve the convergence speed and decrease the final misadjustment error more effective compared with the classical NPCA algorithms and the fixed momentum factor NPCA algorithms.