It is necessary but difficult to obtain a large number of labeled samples to train the classification model in many real scenes. This letter proposes an approach for specific emitter identification(SEI) by introducing model-agnostic meta-learning, which can achieve high accuracy in the case of a limited number of labeled training samples. Specially, we improve the approach to make it suitable for the classification of electromagnetic signals of multiple types of equipments, without spending a lot of time and data to retrain the model structure. The data collected from ZigBee devices and UAVs are used to verify the proposed approach. The simulation results shows that the accuracy of proposed approach can reach more than 90% even though the training task and testing task are two types of devices.

This paper investigates the effect of hardware impairments on the physical layer security for a cell-free massive multiple-input multiple-output (MIMO) network in the presence of pilot spoofing attack. By employing a classical additive hardware distortion model, the joint hardware impairment effects brought by both access points (APs) and user equipments have been taken into account, whereas the eavesdropper is assumed to be equipped with perfect hardware. Thereby, we derive a closed-form lower bound for the ergodic secrecy rate in the presence of imperfect channel state information. To obtain further insights, we investigate asymptotic secrecy performance under different hardware scaling factors. It proves that the hardware-quality scaling effect vanishes as the number of APs increases in the considered secure cell-free massive MIMO system with active attack. Furthermore, by using continuous approximation and path-following algorithms, the optimal power allocation scheme is obtained to maximize the achievable secrecy rate. Numerical results validate the derived results and the efficiency of the proposed power allocation scheme.

Spectrum sensing is one of the crucial technologies used to solve the shortage of spectrum resources. In this letter, based on the short-time Fourier transform (STFT) and convolutional neural network (CNN), we firstly develop a STFT-CNN method for spectrum sensing. The proposed method exploits the time-frequency domain information of the signal samples and achieves the state of the art detection performance. In particular, the method is suitable for various primary users’ signals and does not need any priori information. Besides, we also analyze the signal-to-noise ratio robustness and the generalization ability of the proposed algorithm. Finally, simulation results demonstrate that the proposed method outperforms other popular spectrum sensing methods. Notably, the proposed method can achieve a detection probability of 90.2% with a false alarm probability of 10% at SNR = -15dB.

Due to the emergence of non-terrestrial platforms with extensive coverage, flexible deployment, and reconfigurable characteristics, the hybrid satellite-aerial-terrestrial networks (HSATNs) can accommodate a great variety of wireless access services in different applications. To effectively reduce the transmission latency and facilitate the frequent update of files with improved spectrum efficiency, we investigate the performance of cache-enabled HSATN, where the user retrieves the required content files from the cache-enabled aerial node (AN) or the satellite with the non-orthogonal multiple access (NOMA) scheme. If the required content files of the user are cached in the AN, the cache-enabled node would serve directly. Otherwise, the user would retrieve the content file from the satellite system, where the satellite system seeks opportunities for proactive content pushing to ANs during the user content delivery phase. Specifically, taking into account the uncertainty of the number and location of ANs, along with the channel fading of terrestrial users, the outage probability and hit probability of the considered network are, respectively, derived based on stochastic geometry. Numerical results unveil the effectiveness of the cache-enabled HSATN with the NOMA scheme and proclaim the influence of key factors on the system performance. The realistic, tractable, and expandable framework, as well as associated methodology, provide both useful guidance and a solid foundation for evolved networks with advanced configurations in the performance of cache-enabled HSATN.

Deep cooperative sensing is a cooperative spectrum sensing (CSS) algorithm based on a deep neural network (DNN). Since training DNN requires a large amount of sample data, what is worse existing algorithms directly send training data to the fusion center (FC), which greatly occupies the transmission channel. Motivated by this, in this letter, we introduce the federated learning framework to CSS and propose a federated learning-based spectrum sensing (FLSS) algorithm. In the federated learning framework, there is no need to gather data together. Each secondary user (SU) uses local data to train the neural network model and sends the gradient to FC to integrate the parameters. This framework can perform collaborative training while ensuring local data privacy and greatly reducing the traffic load between SU and FC. Besides, we adopt an efficient network model ShufflenetV2 to reduce the number of parameters and improve training efficiency. Simulation results demonstrate that the FLSS can achieve a detection probability of 98.78% with a false alarm probability of 1% at SNR = −15dB.

Constructing an infrastructure for spectrum sensing and spectrum sharing on the cloud is a promising technology. In this paper, we propose an analysis framework based on spectrum data gathered by distributed sensors to construct the radio environment map (REM). To the best of our knowledge, this is the first attempt to leverage the power of Bayesian analysis and Markov chain Monte Carlo (MCMC) in REM. Specifically, a three-stage Bayesian hierarchical model (BHM) is established to imitate the spectrum data generation process under spatially correlated shadow fading. Parameters of BHM are estimated with the MCMC algorithm from data collected by the sensor network. Then, we address the space-dimension spectrum inference problem with the aim to interpolate the signal strength where there is no sensor node by composition sampling. At each point in the area of interest, the posterior predictive distribution of the receive signal strength can be obtained by kernel density smoother. We make spatial inference under two sensor location modes (square lattice located sensors and randomly located sensors) and two scenarios (with and without the information about the signal source), respectively. Simulation results demonstrate that although the randomly located sensors mode is suitable for parameter estimation, the inference performance is not better than the square lattice located sensors mode. Quantitative analysis of the inference performance confirms that the effectiveness of our data analysis framework is compelling.