最新刊期
- 摘要：To reduce uplink packet delay and energy consumption under channel fluctuations and finite-blocklength transmission errors, a stochastic service model incorporating ordered gating and decoding reliability was developed for cell-free massive multiple-input multiple-output systems. A statistical delay bound was derived using stochastic network calculus under sufficient conditions including independence between arrivals and service, and was used as a delay reference threshold under dynamic policies. A network calculus aided proximal policy optimization algorithm was proposed to jointly select user equipment–access point associations, bandwidth, and transmit power by minimizing delay excess and wireless transmission energy consumption. Reductions in delay excess and transmission energy consumption were observed in simulations. The proposed method provides a reference for uplink association and resource optimization subject to statistical delay constraints.关键词：cell-free uplink;stochastic network calculus;statistical delay constraints;proximal policy optimization3|0|0更新时间：2026-10-12
- 摘要：To address spatially non-stationary (SnS) channel estimation in terahertz (THz) ultra-massive MIMO systems, where hybrid-field path amplitudes vary continuously across antenna elements, a rank-matching-based amplitude attenuation factor (AAF) generation method was extended to array-of-subarrays (AoSA) hybrid-field channels, and an AAF-based SnS hybrid-field estimation scenario was constructed. A gated multi-scale attention (GMSA) network comprising GCIU, MRFU, and SEGA was then developed to replace the nonlinear mapping in PR-DEN, yielding PR-DEN-GMSA. Simulations showed that its five-point average NMSE over three independent runs was −17.076±0.030 dB, improving PR-DEN by 0.696 dB. Under parameter- and computation-matched settings, NMSE was improved by 1.023 dB and 0.816 dB, while the parameter count and per-call theoretical FLOPs were reduced by 41.6% and 22.1%. The proposed algorithm remains stable under the tested variations in AAF correlation length, inter-subarray spacing, and absorption coefficient, and achieves a smaller maximum NMSE degradation under severe undersampling.关键词：terahertz communication;ultra-massive MIMO;spatial non-stationarity;model-driven iteration9|1|0更新时间：2026-10-10
- 摘要：Copyright protection for generative images now extends beyond final outputs to training data, model parameters, generation processes, and provenance credentials. The evolution of generation paradigms further changes where copyright identifiers can be embedded and under what conditions evidence can be verified. Existing surveys mainly focus on specific watermarking mechanisms or attack and defense processes, while the relationships among generation paradigms, protected objects, and evidence goals remain insufficiently compared. To address this gap, studies from 2023 to 2026 were reviewed from three perspectives: generation paradigms, copyright objects, and evidence mechanisms, with relevant earlier work included for context. Embeddable variables and attack surfaces were compared across generative adversarial networks, latent diffusion models, visual autoregressive models, and hybrid architectures. The mechanisms and applicability of proactive watermarking, training-data and model ownership protection, passive forensics, and external content credentials were examined. Results reported in public benchmarks were analyzed to examine trade-offs between watermark detection performance and image fidelity and to discuss tamper localization performance, while distinguishing shared test conditions from differences in method configurations. A small-scale empirical study was used to illustrate the conditions for calculating the joint retention rate of evidence from multiple sources. Existing methods remain limited under generative editing, cross-paradigm transfer, and misattribution attacks; performance on an individual detection or localization task is insufficient to establish the validity of a complete evidence set. Finally, future research directions were discussed in terms of cross-paradigm adaptation, semantic binding, public verification, and evidence continuity across media.关键词：generative images;digital watermarking;copyright protection;content provenance;training-data copyright;Coalition for Content Provenance and Authenticity (C2PA)16|10|0更新时间：2026-10-09
- 摘要：The Verifiable Timed Signature achieves the delayed disclosure of the signature by introducing a time-delay mechanism, and allows the verifier to verify or restore the signature after meeting the predetermined time conditions. The existing VTS schemes mainly focus on the delayed disclosure of the signature and the enforcement of it, providing effective solutions for signature scenarios with time constraints. However, the traditional VTS lacks the ability to dynamically manage the state of specific signatures that have been submitted but not yet disclosed, making it difficult to meet the requirements for signature revocation in some application scenarios. To address these issues, we formally propose the Revocable Verifiable Timed Signature, provide its formal definition and security model, and construct a specific instantiation scheme. The security analysis shows that this instantiation scheme satisfies the security properties such as correctness, unforgeability, soundness, privacy, policy soundness, and policy completeness under the defined security model. Based on this, we abstract the revocation behavior as a computable policy function, adjusts the definition of the policy function to achieve different revocation strategies, and provides corresponding application scenarios for different revocation strategies, thereby providing a more flexible signature state management mechanism for timed signatures.关键词：Revocable Verifiable Timed Signature;verifiable timed signature;Verifiable Delay Function11|8|0更新时间：2026-10-09
- 摘要：To address the issues of high visibility, limited generalization, and the trade-off between attack effectiveness and clean performance in lightweight multi-label image classification, a backdoor attack method was proposed by integrating a content-adaptive deep steganographic trigger with multi-label conditional generation. A shallow convolution-based trigger encoder was used to generate amplitude-constrained pixel-level residuals according to input content, and a multi-label conditional StyleGAN2-ADA was introduced to generate auxiliary samples based on real label distributions. Real and generated triggered samples were jointly optimized with the classifier and trigger encoder. Experiments on two datasets achieved target attack success rates of 98.21% and 99.47%, with mAP decreases of only 0.48 and 0.43 percentage points and false activation rates of 3.96% and 3.12%, respectively. The triggered images achieved PSNR values of 40.76 and 41.18 dB, SSIM values of 0.9928 and 0.9934, and LPIPS values of 0.014 and 0.013. Heterogeneous transfer experiments maintained attack success rates above 93.9%. The proposed method provides an effective and stealthy backdoor attack framework for lightweight multi-label image classification.关键词：Deep steganography;backdoor attacks;multi-label image classification;StyleGAN2-ADA13|7|0更新时间：2026-10-09
-
Reconfigurable PASS-based air-ground 3D coverage optimization for low-altitude intelligent networks AI导读
摘要：To address the difficulty of conventional fixed antennas in simultaneously ensuring satisfactory coverage performance for both aerial users (AUEs) and ground users (GUEs) in low-altitude intelligent networks, a three-dimensional (3D) coverage optimization method based on a reconfigurable waveguide-fed pinching-antenna (PA) system is proposed. First, by jointly optimizing PA positions along the waveguides and the waveguide power allocation, a coverage optimization model is formulated to maximize the weighted spectral efficiency of AUEs and GUEs, subject to the total system power constraint, the minimum signal-to-interference-plus-noise ratio (SINR) constraints of users, and the PA geometric constraints. Second, to address the non-convexity and non-smoothness caused by the discrete association between users and waveguides, a soft association mechanism is introduced to obtain a smooth approximation of the optimization problem, while a variable-shift transformation is employed to effectively handle the PA geometric constraints. Finally, a joint optimization algorithm for PA positioning and waveguide power allocation is developed based on the primal–dual projected gradient method. Simulation results demonstrate that the proposed algorithm can significantly improve the SINR and spectral efficiency of both AUEs and GUEs, thereby achieving balanced 3D coverage enhancement.关键词：Pinching antenna system;PA position optimization;waveguide power allocation;3D coverage optimization22|7|0更新时间：2026-10-09 -
A Lightweight UAV RF Identification Method Based on Transient Envelope Modulation Spectrogram Images AI导读
摘要：In response to the requirements of accuracy, real-time performance, and edge inference efficiency for identity recognition in non-cooperative low-altitude UAV surveillance, this paper proposes a lightweight RF identity recognition method based on Transient Envelope Modulation Spectrogram Images (TEMSI). To address the problem that steady-state payload components, environmental noise, and channel impairments in the complete RF signal obscure hardware transient features, a second-order Haar wavelet is employed for signal denoising. In addition, transient segments are adaptively localized by combining short-time Fourier transform (STFT), energy-trajectory analysis, and sample-level global adaptive threshold detection. Furthermore, a modulation spectrum is extracted from the transient energy envelope using Welch spectral estimation, and a two-dimensional TEMSI representation is constructed to improve discriminability across different UAV models and operating states. Then, by leveraging transfer learning and hierarchical fine-tuning, a lightweight convolutional network is used to accomplish “model–state” RF identity recognition for ten classes of UAVs. Experimental results on the DroneRF dataset show that EfficientNet-B0 achieves both 99.99% recognition accuracy and F1 score, while SqueezeNet requires 10.34 ms for single-sample inference in a general CPU environment. The results indicate that the proposed method maintains high recognition accuracy while effectively reducing inference overhead, providing a reference for edge-side deployment of RF-based UAV identity recognition.关键词：unmanned aerial vehicle;device identification;radio-frequency fingerprinting;RF transient;transient envelope modulation spectrogram image;lightweight neural network17|5|0更新时间：2026-10-09 - 摘要：Searchable encryption protocols enable users to perform keyword searches on encrypted data in the cloud without revealing plaintext, making them a key technology for securing cloud-stored data. However, most existing schemes leak access patterns during query execution—specifically, the sets of matching documents returned by each query and their corresponding relationships. Recent studies have shown that attackers can exploit such leaked access patterns to launch query reconstruction, database reconstruction, and even file injection attacks. This paper provides a systematic review of access pattern protection techniques in searchable encryption. First, we establish a leakage lineage model centered on access patterns, characterizing the hierarchical relationships and combined exploitation among core leakage types—including access patterns, search patterns, and volume patterns. Second, we propose a four-dimensional technical classification framework, categorizing and conducting in-depth comparisons of representative schemes along fourresearchdirections: ORAM/OMAP, obfuscation and differential privacy, distributed multi-server architectures, and关键词：searchable encryption protocol;access pattern protection;oblivious random access machine (ORAM);forward privacy;backward privacy15|5|0更新时间：2026-10-09
-
- 摘要：Deep learning-driven super-resolution reconstruction of precipitation data has achieved significant progress. However, with the rapid proliferation of derivative models, copyright protection and provenance verification of precipitation super-resolution models have become increasingly challenging. Accurately determining the derivation relationship between a candidate model and a protected source model is therefore crucial for safeguarding the intellectual property of precipitation super-resolution models. Nevertheless, super-resolution models produce continuous spatial fields and thus lack the explicit decision boundaries provided by discrete labels in classification tasks. Meanwhile, independently trained models developed for the same task may also exhibit similar behaviors due to shared task-specific regularities and similar data distributions. To address these challenges, this paper proposes a black-box fingerprint verification method for precipitation super-resolution models based on source-behavior deviation (SBD). First, to address the absence of explicit decision boundaries in continuous-field outputs, a locally conservative counterfactual paired-query method is proposed. By bidirectionally redistributing local precipitation amounts under a conservation constraint, the method elicits model responses and constructs local behavioral representations using only the final predictions of the model. Second, to mitigate the interference of behavioral similarity among independently trained models in derivation identification, an SBD method based on a source-behavior reference subspace is developed. Specifically, the variation space defined by the behavioral responses of the source model is used as a unified reference, and the behavioral deformation of a candidate model relative to the source model is projected onto the orthogonal complement of the source-behavior reference subspace. The normalized Frobenius norm is then employed to quantify the degree of orthogonal deviation that cannot be explained by the behavioral structure of the source model, thereby enabling effective verification of derivation relationships under a strict black-box setting. Experimental results on 28 candidate models show that, with a query budget of
, SBD correctly identifies 15 source-derived models and rejects all 10 independently trained models, achieving an accuracy of 89.29%, a balanced accuracy of 91.7%. Relying solely on the final model outputs, SBD can effectively distinguish genuine derivation relationships from behavioral similarities induced by shared task characteristics under strict black-box conditions. 关键词：black-box fingerprint verification;model fingerprint;precipitation super-resolution;source-behavior deviation;counterfactual perturbation37|61|0更新时间：2026-09-24 - 摘要：With the development of millimeter-wave technology, ISAC systems offer high resolution and wide bandwidth but suffer from limited RF chains and high energy consumption. To address these challenges, this paper proposes a dynamic subarray hybrid beamforming scheme for joint communication, sensing, and energy efficiency optimization. A multi-user multi-target millimeter-wave ISAC system with a partially connected hybrid architecture is first established. An energy-efficiency maximization problem is then formulated under communication and sensing constraints, where a low-complexity sensing surrogate metric based on the structure of the FIM is adopted to characterize sensing performance. To tackle the coupling among subarray partitioning, analog precoding, and digital precoding, as well as the non-convex constant-modulus constraints, a two-stage optimization framework is developed. Specifically, a fixed subarray structure is first designed, followed by greedy dynamic subarray partitioning based on the user-weighted total gain. The digital precoder is then optimized using successive convex approximation and semidefinite relaxation. Simulation results show that the proposed scheme significantly reduces the number of RF chains and hardware complexity while achieving performance close to fully digital beamforming and improving energy efficiency in multi-user multi-target scenarios.关键词：integrated sensing and communications;hybrid beamforming;energy efficiency;millimeter wave41|35|0更新时间：2026-09-24
- 摘要：In multi-unmanned-aerial-vehicle (UAV)-assisted Internet of Vehicles (IoV) systems, dynamic variations in traffic flows and task loads can cause mismatches between communication and computing resource supply and demand. To support delay-sensitive tasks, a coordinated resource scheduling method based on a flow matching model was proposed. An optimization model considering task delay, system energy consumption, and deadline violations was formulated, and a three-layer trajectory–resource–offloading mechanism was designed. UAV deployment was optimized through convex inner approximation and sequential optimization, resource-priority and transmit-power candidates were generated and screened using the flow matching model, and task partitioning was optimized based on path availability and queueing feedback. In simulations, lower total system delay and energy-delay product, as well as a higher task success rate, were achieved by the proposed method. The proposed method effectively improves service assurance for delay-sensitive tasks.关键词：IoV;UAV-assisted edge computing;flow matching model;coordinated resource scheduling;service assurance52|41|0更新时间：2026-09-20
- 摘要：Multimodal recommender systems are widely used in content delivery and online services. The third-party visual and textual pretrained encoders are often reused to downstream recommenders through feature caches. This decoupled deployment improves efficiency but exposes pretrained components and cached representations to supply-chain backdoors. If encoder outputs or cached features are covertly poisoned, attackers can bias target-user rankings without changing model parameters or training data, while keeping overall performance nearly unchanged. To address this threat, we propose MAGuard, a multi-view anomaly-gated defense framework. MAGuard detects poisoned representations by fusing local consistency, collaborative behavior, ranking sensitivity, and global distributional deviation into an item-level risk score. It then repairs recommendations at inference time by gating untrusted multimodal features and calibrating collaborative filtering scores, without model retraining. Experiments on MovieLens-10K and Amazon-TG show that MAGuard effectively detects embedding-level backdoors and mitigates attacks with minimal performance loss. This work supports pretrained-component auditing and AI supply-chain protection for multimodal recommender systems.关键词：Multimodal recommendation system;pre-trained encoder;supply chain security;backdoor defense65|59|0更新时间：2026-09-20
- 摘要：Existing glaucoma screening methods based on Convolutional Neural Networks (CNNs) suffer from weak global modeling capability and insensitivity to positional information, owing to the limitation of local receptive fields. Meanwhile, Vision Transformers suffer from high computational complexity and are prone to overfitting on small- and medium-scale datasets. To address these issues, this paper proposes a quantum theory-inspired method for glaucoma screening. The core of the proposed method is a quantum-inspired dynamic anchor bidirectional feature interaction mechanism, which dynamically generates complex-valued anchors, performs joint phase-amplitude modulation, and conducts canonical connectivity-driven feature aggregation. This mechanism achieves efficient local-global feature fusion while reducing computational complexity. In addition, a global context enhancement module and a self-interaction module are introduced to effectively improve the localization of key features and the generalization ability for scale variations. Experiments on public datasets such as AIROGS demonstrate that the proposed method achieves relatively good performance in glaucoma screening with fewer parameters and lower computational cost.关键词：Glaucoma screening;Quantum theory;Dynamic anchors;feature interaction;feature fusion52|45|0更新时间：2026-09-18
- 摘要：Temporal knowledge graph completion aims to predict future missing entity links based on historical observed facts. To overcome the issues of remote information truncation and redundant noise, a method named Historical Summary Enhancement (HSE) was proposed. In this method, entity global historical behavior representations were first extracted through a relation graph neural network. Subsequently, compact historical summaries were constructed to condense cross-temporal historical information. Finally, the historical summaries and current query representations were adaptively combined via a gated fusion mechanism for link prediction. Better results were achieved by the proposed method on the ICEWS14 and ICEWS18 datasets, with the MRR and Hits@10 on ICEWS14 reaching 47.41% and 71.00%, respectively. This effectively improves the performance of temporal knowledge graph completion.关键词：Temporal knowledge graph completion;Historical summary enhancement;Long-range temporal dependency;Time encoding;link prediction37|25|0更新时间：2026-09-18
-
Underwater acoustic ISAC beamforming based on deep reinforcement learning with mixture of experts AI导读
摘要：To achieve the trade-off between communication and sensing performance in underwater acoustic integrated sensing and communication systems, an efficient and highly generalizable beamforming scheme was proposed. First, to address the issues of outdated channel state information and cross-time slot echo interference caused by long propagation delays, a deep reinforcement learning framework was introduced for beamforming decision-making. Second, to overcome the limitations of traditional deep reinforcement learning methods in adapting to diverse performance requirements, a neural network based on a multi-task mixture-of-experts model was constructed to adapt to the trade-offs between communication and sensing performance under different objective-preference factors. In addition, to characterize the long-term impact of current beamforming decisions on system performance, a multi-step cross-task experience replay mechanism was designed, which combines multi-task samples and multi-step rewards to improve learning efficiency. Simulation results demonstrated that compared with benchmark schemes, the proposed scheme obtained more Pareto-efficient solutions under different objective-preference factors and exhibited stronger generalization capability.关键词：underwater acoustic networks;integrated sensing and communication;deep reinforcement learning;mixture of experts21|27|0更新时间：2026-09-18 - 摘要：To address the characteristics of complex perception tasks in high-speed railway scenarios, including directed acyclic graph (DAG) dependencies, high-speed mobility, and dynamically varying trackside computing resources, a mobility- and dependency-aware hierarchical proximal policy optimization algorithm (MDH-PPO) was proposed, since conventional offloading methods designed for independent tasks cannot efficiently support online decision-making under subtask dependency constraints. First, complex perception services were modeled as DAG tasks, and a collaborative multi-edge computing model was established to characterize input uploading, inter-server migration, task queueing, and computation. The dynamic offloading problem was then formulated as a Markov decision process. A hierarchical offloading mechanism comprising topology scheduling and server decision-making was developed, together with a joint action mask. An enhanced proximal policy optimization method was further employed to learn an online offloading policy in high-mobility environments. In the simulations, lower average end-to-end latency and higher system computational efficiency were achieved with MDH-PPO than with the benchmark algorithms under varying computing loads, edge computing capacities, and train operating speeds. These results demonstrate the algorithm’s stable training performance and strong adaptability to high-speed mobility.关键词：high-speed railway;edge computing;dependency awareness;task offloading;proximal policy optimization42|32|0更新时间：2026-09-18
- 摘要：The channel estimation problem in high-speed UAV THz UM-MIMO systems under hybrid near-/far-field propagation, wideband beam squint, Doppler spread, and attitude variations was investigated. A wideband time-varying hybrid-field channel model was established, and an FPN-OAMP-DFMN-TWF method was proposed. Adjacent-slot correlation was exploited by TWF, and multi-slot observations were fused using learnable weights. DFMN, the nonlinear estimator in FPN-OAMP, was composed of adaptive multi-scale fusion and cross-block attention modules; multi-scale angular-domain features were extracted and inter-block information interaction was enhanced. Lower NMSE and faster convergence than FPN-OAMP were demonstrated by simulations, and stable performance was maintained under different UAV speeds, path numbers, and undersampling ratios. The proposed method achieves robust channel estimation for high-speed UAV THz UM-MIMO systems.关键词：Terahertz Communications;ultra-massive MIMO;channel estimation;UAV Communications;Hybrid Near-far Field;attention mechanism37|21|0更新时间：2026-09-17
- 摘要：To address the mismatch between TAS-gated transmission and CQF-periodic forwarding caused by dynamic changes in link delay and frequent handovers in low-Earth orbit satellite networks, as well as issues such as sudden increases in delay, out-of-order delivery, and packet loss for time-sensitive services, a handover-aware TAS-CQF cooperative scheduling deterministic guarantee method is proposed. This method jointly adjusts the CQF period offset and TAS transmission phase based on the comprehensive link delay to achieve period alignment under dynamic link conditions. Furthermore, deterministic network calculus is used to obtain the delay boundaries of the old and new paths, quantify the maximum path differential delay, and adaptively configure the MBB dual-path protection interval and reordering buffer accordingly to ensure continuous and orderly transmission of services during handover. Simulation results show that the proposed method can effectively reduce the packet loss rate and maximum end-to-end delay of time-sensitive services, improve effective throughput, maintain good transmission stability under satellite handover and delay disturbance conditions, and reduce redundant protection overhead in handover scenarios.关键词：LEO satellite network;TAS-CQF;Coordinated scheduling;Certainty guarantee;network calculus39|26|0更新时间：2026-09-16
- 摘要：With the rapid growth of AI, data-center energy consumption continues to increase. To balance computing service requirements and renewable-energy fluctuations, a grid carbon-intensity-aware power–computing coordinated scheduling method was proposed. First, a scheduling framework for geographically distributed heterogeneous data centers was developed to characterize how carbon emission attributes on the power supply side influence task allocation on the computing side. Then, considering workflow dependencies, heterogeneous resources, and spatiotemporal variations in grid carbon intensity, a GINEConv-PPO scheduling mechanism was proposed. A temporal planning algorithm extracted critical-chain positions and resource-contention risks to construct temporal-priority features, while GINEConv was integrated with PPO for low-carbon task allocation under QoS constraints. Simulations across workflow types and sizes showed that GINEConv-PPO met job-batch deadlines. Relative to the lowest emissions achieved by the compared algorithms in each scenario, its emission deviation remained near 0%, versus approximately 6%–35% for the baselines.关键词：Power-Computing co-scheduling;task scheduling;carbon emissions;deep reinforcement learning53|50|0更新时间：2026-09-16
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
- Total Visits：654455 Visits Today：815 Online Users：10
0