The cognitive satellite-terrestrial network (CSTN) has emerged as a transformative architecture for enabling ubiquitous global connectivity, yet its broadcast nature and heterogeneous service demands pose critical security challenges against increasingly sophisticated wiretap threats. This paper proposes a novel reconfigurable intelligent surface (RIS)-assisted secure sparse code multiple access (SCMA) framework in CSTN, which aims to maximize the achievable secrecy rate by jointly optimizing the transmit beamforming, RIS reflection matrix, and SCMA codebook configuration while satisfying power constraints at the satellite and base station and meeting the quality-of-service demands of legitimate users. Specifically, a realistic RIS model incorporating coupled phase-amplitude constraints is considered for practical deployment scenarios. To solve the above non-convex optimization problem in dynamic environments, an intelligent decision-making mechanism based on a modified multi-agent two-delay deep deterministic (MMTD3) algorithm is developed to effectively decouple continuous beam control and discrete codebook selection, and provide a new paradigm for AI-driven cross-domain security optimization in CSTN. Simulation experiments demonstrate that the proposed framework outperforms existing benchmarks in key metrics including convergence, reward values, and secrecy rate, highlighting the framework’s potential for supporting ubiquitous security and massive heterogeneous service demands in CSTN.

As a key component of the space-air-ground integrated network (SAGIN), low Earth orbit (LEO) satellites aim to provide global coverage and reliable services under high-speed mobile conditions, which are critically challenged by severe Doppler effect and inherent broadcast security threats. To address these issues, this paper investigates a multi-reconfigurable intelligent surface (RIS)-assisted orthogonal time frequency space (OTFS) downlink transmission system, where a LEO satellite serves multiple information receivers and potential eavesdroppers acting as energy receivers via simultaneous wireless information and power transfer (SWIPT). By jointly optimizing multi-dimensional resource variables, such as transmit beamforming and RIS reflection coefficients of the spatial domain, and the symbol scheduling matrix of the time-frequency domain, this paper aims to maximize the sum secrecy rate while satisfying constraints on satellite transmit power, the legitimate users’ quality of service, and energy-harvesting requirements. Given the high-dimensional, non-convex, and NP-hard nature of this problem, we develop an enhanced actor-critic deep reinforcement learning (DRL) framework. The core innovation lies in designing an episodic return-prioritized experience selection mechanism with online mixing, which significantly improves the sampling efficiency and policy stability by intelligently selecting training data. Simulation results demonstrate that the proposed approach outperforms existing schemes in achieving a higher sum secrecy rate, providing a practical and highly efficient resource scheduling solution for building secure and reliable next-generation SAGIN.

In this paper, a two-layer game based joint time and power allocation method for an underlay Energy Harvesting Cognitive Radio (EH-CR) network is proposed. The method first models the interplay between the Primary User (PU) and the Secondary Users (SUs) as a Stackelberg game and then models the interplay among the SUs as a Supermodel game in the underlay EH-CR network. Later, a coefficient for evaluating fair ness is introduced in order to promote fairness among the SUs. Subsequently, the utility function of the primary network and the utility function of the secondary network are defined based on their individual profits. By maximizing the secondary network's utility function, the Supermodel game's Nash Equilibrium (NE) solution is achieved. Then, by substituting the NE solution of the Supermodel game into the utility function of the primary network and then maximizing the utility function of the primary network, the NE solution of the Stackelberg game is obtained. Finally, a deterministic strategy can be obtained, which is the time coefficient of equalized spectrum sensing and the equalized power allocation scheme instead of a probabilistic strategy. Simulation outcomes demonstrate that, under the condition of maintaining the communication quality of the PU, the PU's revenue when PH0 = 0.8 can be improved by 18.2% and when PH0 = 0.6 can be improved by 13.3% compared with the conventional method.

In the context of future 6G networks, reconfigurable intelligent surfaces (RIS) and non-orthogonal multiple access (NOMA) are emerging as pivotal technologies for enhancing signal quality and eliminating coverage blind spots. This paper addresses the issue of secure and resilient transmission in RISassisted NOMA systems. Specifically, the base station transmits private signals to multiple legitimate users while dealing with the threat of potential eavesdropping. To model this challenge, we optimize the beamforming vectors and the RIS phase-shift matrix to maximize the sum secrecy rate while satisfying the user quality of service (QoS) requirements and the power constraints of the base station. Since the problem involves high-dimensional variables and non-convex objective functions, it is difficult to be solved by traditional optimization methods. Therefore, the twin delayed deep deterministic policy gradient algorithm (TD3) based on deep reinforcement learning (DRL) is proposed in this paper to effectively address the complexity of the original problem. Numerical results show that the proposed scheme exhibits satisfactory performance in improving communication security, transmission efficiency, and resistance to channel errors.

Deep learning-based intelligent perception has become pivotal in enhancing the effectiveness of UAV monitoring systems. However, deploying complex models on resource-constrained UAV swarms presents significant shortcomings: existing approaches either compromise model accuracy to enable lightweight deployment or introduce communication delays through cloud offloading. More critically, they generally overlook the fundamental prerequisite of monitoring tasks: maintaining stable coverage of the target area. To address this issue, we proposes AirInfer, an innovative collaborative UAV inference framework with critical zones coverage. We formulate a joint optimization problem of deep learning model partitioning and UAV swarm deployment, to minimize end-to-end inference latency. This complex problem can be decomposed into two sub-problems: model partitioning and UAV deployment, which can be solved efficiently using dynamic programming and successive convex approximation, respectively. On this basis, an iterative algorithm is devised to provide guarantees of$\epsilon$-local convergence. Theoretical analysis and experimental results demonstrate that AirInfer not only guarantees blind spot-free monitoring but also reduces inference latency by at least 37 % compared to existing solutions, achieving a balance between perception performance and mission reliability.

Federated learning (FL), as an emerging distributed machine learning paradigm, allows multiple participants to collaboratively train machine learning models without disclosing raw data. The high mobility and flexibility of Unmanned Aerial Vehicles (UAVs) can be effectively integrated with FL for distributed intelligent sensing, such as disaster response and agricultural monitoring. However, existing UAV-assisted federated learning approaches rarely take into account the unique characteristics of large-scale UAV swarms, which often encounter two critical challenges: biased model convergence resulting from statistical heterogeneity and inefficient training due to resource constraints. To address these issues, we propose FedUD, a novel federated learning framework for UAV networks that integrates adaptive UAV sampling and device-to-device (D2D) communications. FedUD employs a two-tier optimization strategy: (1) a set of leader UAVs is sampled by the base station to maximize statistical representativeness in each communication round, thereby mitigating bias from non-IID data; (2) A set of follower UAVs is selected for each leader UAV to form D2D clusters, which helps reduce communication overhead. Follower UAVs conduct local training and transmit their model updates to leader UAVs through D2D communications. The leader UAVs then aggregate these updates and forward them to the base station for global aggregation. Based on theoretical convergence analysis, we formulate a joint optimization problem for UAV sampling and D2D communications. This problem can be effectively solved using a submodular maximization-based iterative algorithm. Extensive experiments conducted on physical testbeds and simulations demonstrate that FedUD significantly enhances the stability of model performance and reduces system latency during the training process compared to benchmark methods.

Abstract With the rapid development of communication technology and the increasing demand for coverage refinement in wireless communication networks, the optimization of wireless communication networks is faced with unprecedented challenges. Obtaining the signal distribution map of wireless communication networks efficiently has become a popular area of study in this field. This paper considers a distributed sensing network architecture, a radial basis function neural network is used to process electromagnetic data and optimize the parameters of the random forest model. Then, interpolation processing of incomplete electromagnetic data is achieved by the improved random forest model, based on which a signal distribution map of the wireless communication network is reconstructed. The results indicate that the proposed algorithm yields high interpolation accuracy. The average error between the real signal distribution and the reconstructed signal distribution is 2.7973 dBm when the proportion of sampled nodes is 1%, and the similarity of the reconstructed signal distribution map to the original signal distribution map is good, demonstrating certain application prospects.

Since distributed sensing, storage, and computing are the frontiers for future sixth‐generation (6G) communication systems, user terminal (UT) localization based on received signal strength (RSS) data from wireless sensor networks (WSNs) has received widespread attention because of its low energy consumption and ease of operation. Most of the existing work focused on the single‐source localization problem. However, multiple UT localization is a more realistic problem that has not been well addressed. In this paper, we proposed a novel multiple UT localization scheme. Specifically, based on the log‐normal property of spatial shadowing, the RSS is approximated as a random variable obeying a log‐normal distribution, and the objective function is derived via maximum likelihood estimation. Then, aiming to better solve the objective function, a radio map is constructed to narrow search area, and a meta‐heuristic algorithm with global search capability is adopted. Compared with the state‐of‐the‐art methods through simulation experiments, it is proved that the method proposed in this paper has the best localization performance.

Deep learning-based automatic modulation recognition (AMR) is essential for enabling access authorization and spectrum management for interconnected devices and sensors in Internet of Things (IoT) systems. However, the open collection of data and the use of third-party training resources may introduce security vulnerabilities, especially backdoor attacks. Current research allows attackers to mislead receivers into misclassifying signals as a specific modulation type, but the fixed position of the trigger restricts its use in complex wireless networks. In this work, we propose a novel spatially distributed multi-target backdoor attack (SMBA) method. This method utilizes a trigger pattern to manipulate all types of input signals into multiple target modulation types by embedding the stealthy trigger at different spatial locations within the input signals. SMBA disseminates malicious samples across multiple target types specified by attackers, making it difficult for defenders to predict the target type into which the modulation signal embedded with the trigger is classified. This work reveals new security threats to AMR and provides important insights for developing defense technologies for IoT systems.

Radio frequency fingerprint identification (RFFI) has become a crucial technology in physical layer authentication, and plays an important role in authenticating the identities of wireless communication devices in the Internet of Things (IoT). Although open-set recognition has been applied in RFFI tasks, these schemes still demand extensive RF signal samples. In this paper, few-shot open-set recognition is being dedicated to exploring in RFFI tasks. To surmount mentioned challenges, we propose meta-learning by gaussian prototype network (MLGPN) scheme to achieve the goal of few-shot open-set recognition. MLGPN adopts the Mahalanobis distance between the embedding feature and the gaussian prototype as its metric. With the introduction of open-set loss function, the proposed scheme shows excellent open-set recognition performance. It is worth mentioning that meta-learning not only satisfies the demands of few-shot scenarios, but also enables new devices to join and leave without the need for retraining. Experiments conducted based on real LoRa RF signals confirmed the excellent performance of our proposed scheme for few-shot open-set recognition which surpasses traditional prototypical network model by 5.3% of AUC and 6.7% of ACC under the 1-shot condition. Compared with other schemes, the proposed scheme also demonstrated significant advantages.

A radio environment map (REM) is a visualization method that display electromagnetic properties, such as received signal strength, channel gain, and power spectrum density in combination with geographic information. The map can effectively support modern Internet of Things (IoT) network planning and resource management. A novel REM construction scheme of an arbitrary height and frequency in 3-D space is studied in this article. First, a complex urban environment is considered, where the radiation sources transmit wireless signals in different frequency bands. Then, the construction is sliced into 2-D planes with various elevations to achieve precise and efficient sensing of 3-D space. For near-ground scenarios, preliminary global interpolation based on linear unbiased estimation is first performed to obtain a coarse REM, and then graph neural networks are utilized to further extract the relationships and features of the spatial nodes to improve the construction accuracy. For high-altitude scenarios, a small range of interpolation is carried out on the basis of linear unbiased estimation with the clustering center obtained by clustering the known sensing nodes as the center of the circle. Then the global construction is implemented via domain transformation processing to increase the construction speed. Finally, the 2-D REMs are stacked in sheets according to elevation to form a 3-D REM. The simulation results demonstrate the effectiveness and superiority of the proposed scheme.

radiation source localization (RSL) via received signal strength (RSS) from wireless sensor network has received much attention due to its simplicity and low energy consumption. However, localization in the presence of multiple radiation sources and in complex propagation environments, such as shadow effect is still not well addressed. In this article, we utilize the lognormal property of the shadow effect to approximate RSS as a random variable obeying a lognormal distribution, and construct a multi-RSL model based on maximum likelihood estimation. Thus the model is shadow-resilient. In order to better solve the nonconvex objective function, we construct radio map based on RSS, so as to provide key parameters of number of radiation sources, initial feasible solutions and location constraints for the localization model. Ultimately, we realize high-precision localization of radiation sources. Among them, this is the first time that sparse Gaussian process regression based on variational inference is applied to radio map construction, and the method can greatly reduce computational complexity to better meet the real demand of large area and large data. Both simulation and real-world experiments show that the proposed method can achieve the highest localization accuracy with the lowest computational complexity compared with the state-of-the-art methods. Additionally, Cramer–Rao lower bound (CRLB) is also derived in detail.

PHY-layer spoofing attack is a potential critical issue in wireless network communication security, which could lead to catastrophic consequences for critical mission and applications, especially in Industrial Internet of Things scenarios with enormous number of devices. In this paper, we propose a novel spoofing attack detection scheme exploiting Channel State Information (CSI) phase difference. Firstly, we establish a mapping between CSI phase difference and the location of wireless communication devices to achieve the goal of spoofing attack detection. Due to the stable property of CSI phase difference, we convert CSI phase difference into heatmaps for subsequent training of the neural network model. Then we propose Wasserstein generative adversarial network and Encoder (WGAN-Encoder) deep-learning-based model in the scheme. This model utilizes discriminator feature residual error and image reconstruction error to get anomaly score for spoofing attack detection. This model overcomes the limitations of traditional detection methods on prior knowledge the attacker’s real CSI under real communication scenarios. Finally, we carry out extensive experimental evaluations about the detection performance and robustness of the proposed scheme based on data collected in time-varying scenarios. The results have successfully demonstrated that the proposed scheme exhibits outstanding performance.

Abstract Based on the successful application of generative adversarial network (GAN) models in the field of image generation, this article introduces GANs into the field of deep learning for communication systems and surveys its application in modulation classification. To solve the difficulties in feature extraction, to address the low recognition accuracy of existing radio signal modulation‐type recognition methods, and to adapt to complex electromagnetic environments with high noise interference intensity, this article presents a modulation recognition model for high‐order digital signals. This model uses the Morlet wavelet transform to analyse time‐frequency signals, uses the excellent image generation performance of a GAN model to extract and reconstruct the features of noise‐contaminated time‐frequency images, and designs an integrated classification network architecture to classify and predict reconstructed images. The experimental results show that the algorithm model proposed in this article can significantly improve the recognition accuracy of high‐order digital modulated signals under low signal‐to‐noise ratio conditions and can achieve 90% recognition accuracy at a signal‐to‐noise ratio of 1 dB.

TarGuess − I is a leading online targeted password guessing model using users’ personally identifiable information (PII) proposed at ACM CCS 2016 by Wang et al. It has attracted widespread attention in password security owing to its superior guessing performance. Yet, after analyzing the users’ vulnerable behaviors of using popular passwords and constructing passwords with users’ PII, we find that this model does not take into account popular passwords, keyboard patterns, and the special strings. The special strings are the strings related to users but do not appear in the users’ demographic information. Thus, we propose TarGuess − I + K P X , a modified password guessing model with three semantic methods, including (1) identifying popular passwords by generating top-300 lists from similar websites, (2) recognizing keyboard patterns by relative position, and (3) catching the special strings by extracting continuous characters from user-generated PII. We conduct a series of evaluations on six large-scale real-world leaked password datasets. The experimental results show that our modified model outperforms TarGuess − I by 2.62% within 100 guesses.