Issue Ebookcover storyIssue CoverIssue Contents
Online FirstList of IssuesDocument Types
- Abstract:Task allocation in computing power networks faces challenges such as complex task dependencies, multi-dimensional and heterogeneous resources with uneven distribution, and difficulty in global collaboration among large-scale nodes. Traditional single agent algorithms were difficult to cope with the dynamic resource collaboration issues in wide area environments, and often relied on global state information, while existing multi-agent reinforcement learning algorithms lacked targeted modeling of task dependencies and network topology, resulting in low global collaboration efficiency. To address these issues, a distributed collaborative allocation algorithms of dependency-constrained tasks in computing power networks was proposed. First, complex subtask dependencies were processed through layering, triangular dependency decoupling and multi-resource threshold merging. Then, relation-graph attention network (RGAT) was adopted to encode multi-resource relations and topology features, combined with multi-agent soft actor-critic (MASAC) for distributed optimization. Experimental results show that in multiple scenarios, the proposed algorithms significantly optimizes task completion time and energy efficiency compared to baseline algorithms.Keywords:computing power network;task reconfiguration;relation-graph attention network;multi-agent reinforcement learning;task allocation442|344|0Updated:2026-06-20
Few-shot class-incremental learning for specific emitter identification with orthogonal space constraints AI Introduction
Abstract:To address the challenge of few-shot class-incremental learning (FSCIL) for specific emitter identification (SEI) with scarce samples, and the issue that existing methods lacked explicit geometric constraints, leading to new class embeddings being easily confused with those of old classes, a method for FSCIL based on orthogonal space constraints was proposed. Firstly, a set of mutually orthogonal pseudo-target vectors was introduced as structured prior knowledge, and the lower and upper bounds of their quantity were theoretically derived. Next, a collaborative optimization strategy based on orthogonal pseudo-target vectors was proposed, so as to impose geometric constraints on new class representation directions and guide the feature extractor to reserve expandable representation directions in the embedding space. Finally, a classifier weight calibration strategy was designed to quantify the misclassification risk of new class samples during the decision process, using high-risk samples to enhance the decision boundaries. Experimental results on the ADS-B and Wi-Fi datasets show that, the proposed method outperforms other benchmark methods, especially in the extreme case where only one incremental training sample is available.Keywords:specific emitter identification;few-shot class-incremental learning;orthogonal space constraint;classifier weight calibration422|535|0Updated:2026-06-20Shortwave autonomous detection and localization method based on deep reinforcement learning AI Introduction
Abstract:In the process of real-time shortwave signal detection and localization, problems such as data association and error propagation are frequently observed, by which a mismatch between detection and localization results is easily induced, and the timely output of outcomes is impeded. To address the above issues, an integrated shortwave signal detection and localization method based on multi-agent proximal policy optimization was proposed. By constructing a reinforcement learning environment using shortwave signal time-frequency diagrams, spectral characteristics of shortwave signals were effectively captured. A filtering-window agent with hybrid action space was designed for adaptive signal selection. Furthermore, localization error ellipse probability, signal matching degree, and multi-agent collaboration strategies were incorporated into the reward function design. Optimal policies were explored by the optimized multi-agent proximal policy optimization to achieve autonomous shortwave signal detection and localization. Simulation results demonstrate that compared with baseline algorithms, the proposed method reduces average recognition-localization time by 0.12 s while improving accuracy by 22%, thereby providing a novel solution for autonomous cooperative target detection and localization in complex electromagnetic environments.Keywords:Shortwave signal detection and localization;Multi-Agent Proximal Policy Optimization;reinforcement learning;Hybrid action space;localization error ellipses probability359|406|0Updated:2026-06-20Channel estimation algorithm for OTFS based on basis expansion model with bidirectional LSTM network AI Introduction
Abstract:Accurate channel estimation is critical for the reliable deployment of orthogonal time frequency space (OTFS) systems in 6G high-mobility scenarios. To further improve the channel estimation accuracy in OTFS systems, a channel estimation algorithm integrating the basis expansion model (BEM) and a bidirectional long short-term memory (LSTM) network was proposed. Firstly BEM was used to reformulate the channel estimation problem in the delay-Doppler (DD) domain into a basis coefficient estimation problem. Then a bidirectional LSTM network was introduced for offline training and online prediction of the basis coefficients, and a self-attention (SA) mechanism was incorporated to strengthen global perception of temporal features, enabling high-accuracy estimation and dynamic tracking of the basis coefficients. Building on that, a low-complexity two-stage equalizer was further designed. Channel information was reconstructed from the estimated basis coefficients, multipath effects were mitigated via a single-tap equalizer, and a log likelihood ratio based iterative interference cancellation (LLR-IIC) algorithm was employed in the DD domain to suppress residual interference induced by Doppler spread. Simulation results demonstrate that the proposed algorithm offers significant advantages in channel estimation accuracy, generalization capability, and complexity. When combined with the proposed equalizer, superior bit error rate (BER) performance is achieved while maintaining low complexity.Keywords:channel estimation;OTFS;high-mobility scenario;BEM;bidirectional LSTM;two-stage equalizer385|254|0Updated:2026-06-20- Abstract:To address the low tally-verification efficiency and insufficient support for diverse voting modes identified in electronic voting scenarios, an anonymous and accountable aggregate ring signature algorithm supporting multiple voting modes was proposed. The proposed construction was designed to preserve voter privacy while ensuring the legitimacy of voting operations. Ring signature technology was employed to enable decentralized setup and hide verifier identities, reducing the risk of tracing or manipulation during verification, while an aggregation mechanism was introduced to compress multiple ballots into compact verification objects to improve tally-verification efficiency.To regulate the number of votes cast by the same signer, a linking authority was introduced together with linkable tags, enabling support for both single-vote and multiple-vote settings. Security analysis and experimental results show that the proposed algorithm achieves a good balance between anonymous regulation and efficiency, and reduces the time overhead in the linking phase by about 98.29% on average compared with the k-LRS scheme.Keywords:electronic voting;ring signature;identity hiding;linkability340|270|0Updated:2026-06-20
Graph masked contrastive learning method for insider user behavior anomaly detection AI Introduction
Abstract:To address the limitations of existing insider user behavior anomaly detection methods, including weak spatiotemporal correlation representation, difficulty in identifying stealthy anomalous behaviors, and poor robustness to normal business fluctuations, a graph masked contrastive learning method named TGMC for insider user behavior anomaly detection was proposed. Firstly, multi-source heterogeneous logs were organized into a sequence of heterogeneous graphs using time windows to characterize interaction relationships between users and business resource entities. Then, a hybrid encoder composed of a graph attention network and a gated recurrent unit was designed to extract spatial topological features and temporal evolution features, respectively, enabling deep coupling across spatiotemporal dimensions. A masked reconstruction mechanism was introduced to recover the features and edge structures of masked nodes from the neighborhood context, mining neighborhood correlation patterns to enhance the detection capability for stealthy anomalous behaviors. Finally, a multi-view contrastive learning mechanism was devised: by constructing augmented views, it enforced semantic consistency of the same user’s representations across views while enlarging the representation distance between different users, thereby learning stable behavior representations and improving recognition robustness. Experiments on the CERT r4.2 and CERT r5.2 datasets demonstrate that, compared with baseline methods, TGMC improves true positive rate (TPR) by 3.59% and 1.59%, and reduces FPR by 5.45% and 2.49%.Keywords:graph representation learning;contrastive learning;insider user behavior;robustness179|216|0Updated:2026-06-20- Abstract:To address the issues of heavy reliance on manual expertise, high time consumption, and increasing model complexity in color image steganalysis network design, a two-stage differentiable steganalysis architecture search called STARS was proposed to automatically design efficient network architectures. In view of the critical role of preprocessing layers, the first stage constructed a specialized search space oriented towards color channels. It utilized group convolution and prior knowledge to guide the optimization of noise residual extraction, demonstrating superior high-pass characteristics and search efficiency compared to existing methods. Building upon this foundation, the second stage performed a joint search for the network’s operation types and depth. Experimental results demonstrate that the architectures derived by STARS outperform state-of-the-art models on the ALASKA-v2 dataset, achieving superior detection performance in both spatial and JPEG domains with significantly reduced model size.Keywords:information hiding;image steganalysis;steganography;neural architecture search;convolutional neural network189|144|0Updated:2026-06-20
Near-field wideband cellular codebook design and channel estimation in terahertz extremely large-scale MIMO AI Introduction
Abstract:In terahertz extremely large-scale MIMO systems, new challenges were imposed by the near-field beam split effect on wideband channel estimation. To address this issue, a subcarrier-dependent wideband cellular codebook design scheme was proposed. This scheme designed different codebook sampling structures for each subcarrier, enabling the refinement of codeword without increasing the codebook dimension. Correspondingly, a wideband channel estimation algorithm named multi-dimensional accumulation generalized simultaneous orthogonal matching pursuit (MAGSOMP) was introduced, which enhanced channel estimation accuracy by accumulating multidimensional channel correlations. Simulation results demonstrate that, at a signal-to-noise ratio of 5 dB, the normalized mean square error of channel estimation achieved by the proposed algorithm improves by more than 7 dB compared to existing algorithms.Keywords:terahertz;XL-MIMO;near-field wideband communication;channel estimation;codebook design;beam split effect262|414|0Updated:2026-06-20Robust asynchronous tensor estimation architecture for mmWave hybrid array uplink sensing AI Introduction
Abstract:To address the limited dimensionality of raw observations in millimeter-wave hybrid-array uplink sensing and the time-frequency offsets caused by clock asynchrony, a multi-dimensional parameter estimation and automatic pairing method based on a regularized signal ratio and tensor sparse reconstruction was proposed, and a robust asynchronous tensor estimation architecture was constructed. Virtual fully digital observations were constructed through time-division beam scanning to restore spatial sampling degrees of freedom, and a dominant reference path was extracted from uplink data to form a reference signal. A noise-power-aware regularized signal-ratio model was introduced to suppress asynchronous phase interference and mitigate the issue of heavy-tailed noise amplification. Subsequently, an equivalent third-order space-frequency-time tensor was constructed, and low-complexity estimation and automatic pairing of the angle of arrival (AoA), time delay, and Doppler parameters were achieved via decoupled subspace estimation and tensor sparse reconstruction. Simulation results show that the proposed architecture maintains stable detection and parameter estimation performance under heavy-tailed noise and severe asynchronous conditions, achieves an effective detection rate above 90% at -5 dB, and enables more than 99% of estimates to converge to millimeter-level accuracy (with a distance error of less than 0.02 m and a velocity error of less than 0.01 m/s) in the medium-to-high SNR regime.Keywords:integrated sensing and communication;Hybrid Array;millimeter wave;Uplink Sensing;parameter estimation;tensor algebra248|409|0Updated:2026-06-20MAD-based variable center and variable kernel width MCC underwater acoustic channel estimation algorithm AI Introduction
Abstract:To address the limitations of conventional maximum correntropy criterion (MCC) algorithms in non-Gaussian, non-stationary underwater acoustic channels, specifically their dependency on fixed kernel widths and the vulnerability of center estimation to impulse contamination, a variable center and variable kernel width MCC (VC-VKW-MCC) channel estimation algorithm based on median absolute deviation (MAD) was proposed. The robust property of MAD to impulses was utilized, combined with sliding windows and soft clipping, to achieve the adaptive updating of the filtering center and kernel width, thereby effectively suppressing impulsive noise. The mean-square convergence condition of the method was theoretically derived. Simulation results show that the proposed algorithm obtains more than 70% higher steady-state accuracy than the fixed-kernel-width MCC algorithm, and more than 60% higher accuracy than the Fe-MCC-AKW algorithm.Keywords:Underwater acoustic channel estimation;maximum correntropy criterion;Median absolute deviation;Variable center;Variable kernel width192|153|0Updated:2026-06-20- Abstract:In multi-terminal downlink semantic communication scenarios, employing a uniform semantic compression ratio caused difficulty in semantic decoding for low-computing-power terminals, as well as underutilized computing power resources and insufficiently refined semantic data for high-computing-power terminals. To address this issue under constraints of limited terminal computing power and link bandwidth, a lightweight Swin Transformer-based resource-adaptive semantic compression strategy was proposed. The semantic encoder integrated a gating network and a sparse attention mechanism to customize differentiated semantic compression ratios for individual terminals. A joint computing and bandwidth allocation model was formulated to minimize total system energy consumption. The proximal policy optimization (PPO) algorithm was employed to solve the optimization problem. Simulation results demonstrate that, compared with fixed compression ratio schemes, the proposed strategy reduces total system energy consumption by 39%.Keywords:semantic communication;Swin Transformer;sparse attention mechanism;proximal policy optimization algorithm287|198|0Updated:2026-06-20
- Abstract:To address the limitation of semantic reconstruction performance in traditional covert communications under low signal-to-noise ratio (SNR) conditions, a semantic covert communication scheme integrating reconfigurable intelligent surfaces (RIS) and a tripartite generative adversarial network (GAN) was proposed. Firstly, the physical layer and the semantic layer were jointly modeled, and an optimization problem was formulated to maximize semantic covert fide-lity. Then, a hierarchical optimization framework was developed. At the physical layer, the transmit power and RIS phase shifts were jointly optimized under Kullback-Leibler (KL) divergence-based covert constraints to construct a channel environment favorable for signal transmission. Based on this design, a semantic-oriented tripartite adversarial training mechanism was introduced at the semantic layer, and end-to-end learning was enabled to jointly optimize semantic reco-very performance and covertness. Experimental results show that, the proposed scheme achieves a 29.8% improvement in BLEU score compared with the best benchmark scheme, while the eavesdropper’s detection accuracy remains close to 50%.Keywords:covert communication;semantic communication;reconfigurable intelligent surface;generative adversarial network;KL divergence377|532|0Updated:2026-06-20
Meta mean-field Q-learning-based anti-jamming and covert communication mechanism for large-scale UAV networks AI Introduction
Abstract:In dynamic and complex electromagnetic environments where malicious jammers and passive eavesdroppers coexist, large-scale UAV networks impose increasing demand on anti-jamming capability and communication covertness. Based on this, a meta-reinforcement-learning-based anti-jamming covert communication mechanism for large-scale multi-UAV networks was proposed. Firstly, a time-slotted spectrum-sensing model was designed for the multi-UAV network to support the communication process of UAV networks under jamming and eavesdropping. Secondly, a mean-field Q-learning (MFQ) algorithm was designed to achieve distributed channel-access policy learning in large-scale UAV scenarios. Finally, a meta-MFQ (MMFQ) algorithm was proposed to enhance the generalization performance of the channel-access policy to different jamming and eavesdropping scenarios, thereby jointly optimizing anti-jamming performance and covertness. Simulation results demonstrate that, compared with baselines, the proposed algorithm achieves average improvements of 15.41% in normalized reward and 44.56% in convergence speed over large-scale UAV communication scenarios.Keywords:large-scale unmanned aerial vehicle;spectrum sensing;anti-jamming communication;meta reinforcement learning;covert communication349|493|0Updated:2026-06-20- Abstract:In orthogonal time-frequency space (OTFS) systems with zero-Doppler pilot patterns, pilots and data symbols were non-orthogonal in the time-frequency (TF) domain. Moreover, the delay-domain path discrimination approach originally proposed for it from an error floor under limited bandwidth. To achieve simultaneous orthogonality in both domains, the delay dimension of data symbols was compressed, converted to low-dimensional TF symbols via inverse symplectic finite Fourier transform, and mapped to pilot-free indices for orthogonal isolation. To address increased complexity from dimension changes, modulation and demodulation were decomposed into data mapping and conventional processes. For improved robustness in channel estimation, after Bayesian inference, the sparsity of the posterior mean was evaluated to check convergence. If not converged, hyperparameter initial values were perturbed and inference was re-run. Simulation results show that error floors are eliminated by the proposed method, normalized mean square error is reduced by up to 3 dB, required SNR is lowered by up to 7 dB at BER 10-4, and the computational complexity is reduced to about
of direct computation. A physical verification platform using software-defined radio and channel simulator confirms the feasibility and effectiveness of the proposed approaches.Keywords:OTFS;pilot pattern;Bayesian inference;channel estimation222|165|0Updated:2026-06-20 Multi-objective joint scheduling optimization algorithm for time-sensitive networking AI Introduction
Abstract:To address the inadequate optimization of performance indicators, such as end-to-end delay, jitter, and bandwidth utilization, as well as the poor scalability in the joint routing and scheduling of time-sensitive networking (TSN), a multi-objective joint scheduling optimization algorithm based on an improved non-dominated sorting genetic algorithm was proposed. Specifically, a multi-objective network architecture utilizing TSN performance indicators as optimization objectives was established, feasible traffic paths were comprehensively analyzed, and the global optimal multi-objective solution was identified. Furthermore, the effectiveness of the proposed algorithm was evaluated through extensive experiments. The proposed algorithm significantly improves the comprehensive traffic performance in terms of end-to-end delay, jitter, and bandwidth utilization in large-scale complex scenarios.Keywords:time-sensitive networking;non-dominated sorting genetic algorithm;multi-objective optimization;joint routing and scheduling229|163|0Updated:2026-06-20Lightweight malware classification method based on heterogeneous feature collaboration AI Introduction
Abstract:To address the limited information coverage of single-feature static malware classification and the high cost of multi-feature fusion models, a lightweight collaborative method for resource-constrained scenarios was proposed. The method integrated opcode sequences and static application programming interface (API) calls to characterize malware from instruction execution and system interaction perspectives. A multi-scale residual module was used in the opcode branch to capture local patterns and long-range dependencies, while a triplet semantic chain was introduced in the API branch to convert discrete calls into structured functional information, reducing sparsity and redundancy. A position-aware multi-head self-attention mechanism was then employed to dynamically fuse the two features for joint modeling, achieving high accuracy with low model complexity. Experiments on the BIG2015 and enterprise-level PE datasets achieve accuracies of 99.53% and 99.4%, respectively, with only 30.2 MFLOP, demonstrating a high-accuracy and low-overhead solution for malware classification in resource-constrained environments.Keywords:malware classification;deep learning;lightweight;opcode;API167|134|0Updated:2026-06-20
Papers
Survey on spatio-electromagnetic programmable antenna: fundamentals, recent advances, and future directions AI Introduction
Abstract:The movable antenna (MA) technology, which enables manipulation of electromagnetic (EM) radiation characteristics through adjustable antenna positions and postures, has attracted significant attention from both academia and industry in recent years. However, existing interpretations of MA often attribute the reconfigurability of EM radiation characteristics solely to changes in antenna position, while in fact, active reconfiguration at the electromagnetic level is the key factor determining the operational posture of MA. Based on this, the core concept of MA was clarified, emphasizing that electromagnetic reconfigurability was essential for breaking through current communication performance bottlenecks. On this basis, the spatio-electromagnetic programmable antenna (SEPA) was proposed to more comprehensively characterize the synergistic reconfigurability of antennas in both spatial and electromagnetic domains. Finally, several future research directions for SEPA were outlined, including intelligent paradigms for EM radiation reconfiguration, compensation for spatial coverage gaps, platform design, and validation of key technologies.Keywords:movable antenna;SEPA;electromagnetic radiation characteristic;degree of freedom256|124|0Updated:2026-06-20- Abstract:Low earth orbit (LEO) satellite communication network is an important means to address coverage in remote areas and emergency communications. Constrained by the practical limitation of being unable to deploy ground gateway stations globally, China’s LEO satellite networks urgently need to reduce their dependence on terrestrial nodes. Computing and network convergence (CNC) technology, which leverages inter-satellite links and heterogeneous computing power to enable service pre-positioning, is a key approach to breaking through this bottleneck. Firstly, the development status of LEO satellite CNC was reviewed. Then, the main challenges faced by this technology were analyzed. Subsequently, driven by typical services, an “edge-end” bi-level CNC architecture for LEO satellites was designed. Finally, the development trends of this technology were discussed and potential future research directions were outlined.Keywords:satellite communication;low earth orbit satellite;computing and network convergence;edge-end collaboration;service-oriented197|101|0Updated:2026-06-20
Comprehensive Reviews
Improved QAOA based on multi-angle and RY-assisted structure AI Introduction
Abstract:To address the limited expressive capability of the standard quantum approximate optimization algorithm (QAOA) in shallow circuits and the increased noise sensitivity caused by deeper quantum circuits, an improved QAOA, termed MA-RY-QAOA, was proposed by integrating parameter multi-angleization and an RY-assisted structure. In the proposed algorithm, edge-parameter multi-angleization and RY expansion were incorporated into the problem layer, while qubit-parameter multi-angleization was introduced into the mixing layer. Simulation experiments were conducted on the Max-cut problem of random graphs, complete graphs, 2-regular graphs, and 4-regular graphs with circuit depth ( ) and node sizes ranging from 4 to 12. The average approximation ratio was adopted as the evaluation metric. The results showed that the average approximation ratios achieved by MA-RY-QAOA and MA-QAOA were generally superior to those obtained by the standard QAOA, RY-QAOA, and QAOA+ algorithms. In particular, MA-RY-QAOA exhibited better performance on 2-regular and 4-regular graphs and achieved higher approximation ratios for certain scales of random graphs. These findings indicate that MA-RY-QAOA improves optimization performance while maintaining a shallow circuit structure and provides a useful reference for the design of QAOA variants on noisy intermediate-scale quantum (NISQ) devices.Keywords:quantum approximate optimization algorithm;Parameter Multangularization;RY-assisted Structure;Maximum Cut Problem;noisy intermediate-scale quantum device241|355|0Updated:2026-06-20- Abstract:Aiming at the resource scheduling challenges of 5G-R/the future railway mobile communication system(FRMCS) network slicing under service heterogeneity, high-speed node mobility and insufficient incentives, a collaborative framework integrating hierarchical federated learning (HFL), Stackelberg game and multi-task proximal policy optimization (MT-PPO) was proposed. A three-layer HFL was designed to handle non-independent and identically distributed (Non-IID) data and predict resource states. A two-layer Stackelberg game was adopted to improve node participation and coordination efficiency. MT-PPO was employed to fuse both outputs for differentiated scheduling. Simulation results show that the proposed framework outperforms baseline methods in quality of service (QoS), resource utilization and throughput.Keywords:5G-R;FRMCS;network slicing;resource allocation;Multi-Task Proximal Policy Optimization213|188|0Updated:2026-06-20
Correspondences








