To address the susceptibility of audio-based voice interaction to noise under low signal-to-noise ratio conditions and the privacy concerns of vision-based approaches, the recognition of speech commands based on millimeter wave radar sensing vocal-cord vibration was studied. To overcome the challenges posed by weak radar vibration signals and interference from surrounding objects, a multi-antenna beamforming-based signal enhancement method was developed, together with a frequency-domain energy analysis mechanism that removed abnormal components caused by moving objects, thereby extracting stable vibration features. To improve cross-user recognition performance, a few-shot transfer framework, mmFSNet, was designed to adapt the classifier with a small number of target-user samples, enhancing the system’s generalization capability. The system yields an average accuracy of 97.3% over 10 users and 11 speech command categories, with performance variations kept within 5% under changes in distance, angle, and environment. When fine-tuned with only three samples from a new user, the recognition accuracy exceeds 80%. The results demonstrate that the proposed method ensures robust speech command recognition in complex environments and provides effective adaptability to unseen users.
In data-driven electricity theft detection methods, feature extraction plays a critical role in detection accuracy. To address this issue, a novel approach was proposed by integrating a temporal 2D variation modeling network (TimesNet) with a long short-term memory (LSTM) network to fuse time-frequency features. In this method, long-term temporal dependencies in the time domain and periodic patterns in the frequency domain were jointly modeled. An optimization strategy combining class-weighted focal loss was introduced to effectively mitigate the adverse impact of class imbalance on model performance. Furthermore, a multi-scale InceptionBlock combined with an LSTM architecture was designed to enhance the model’s capacity to capture complex electricity consumption patterns. Experimental validation on public datasets demonstrated that the proposed method achieved an F1-score of 0.90 and an area under the curve-receiver operating characteristic (AUC-ROC) of 0.97, representing improvements of 10% and 5%, respectively, over the classical CNN-LSTM model. The results indicate that the proposed method effectively balances low false alarm rates with high detection rates.
With the rapid evolution of 6G wireless communication technologies, smart wearable devices, empowered by ultra-low latency and high-speed connectivity, have been widely deployed, continuously enhancing users' quality of life. However, real-time data interactions over public channels expose the security and privacy of user data to severe threats. To address this challenge, an efficient anonymous authentication protocol tailored for 6G-enabled smart wearable scenarios was proposed. By integrating cryptographic hash functions with physical unclonable functions (PUF), the proposed protocol achieved mutual authentication among smart wearable devices, users, and the remote server, while ensuring the secure generation of session keys. Security analysis demonstrated that the protocol exhibits strong robustness against various known attacks. Furthermore, performance comparisons indicated that, compared with existing representative schemes, the proposed protocol provided more comprehensive security properties and functional support while maintaining low overhead.
With the advancement of satellite remote sensing data acquisition technologies, satellite remote sensing images have experienced explosive growth. However, existing satellite-ground communication technologies are constrained by channel bandwidth, making it challenging to achieve efficient transmission of massive remote sensing images. To address this issue, an adaptive coding and decoding method tailored for satellite remote sensing image transmission was proposed. This method comprised five modules: a global-local feature extraction module, a key semantic feature selection module, a joint source-channel coding module, a joint source-channel decoding module, and an analysis module. Specifically, the global-local feature extraction module leveraged multi-directional Mamba blocks and residual blocks to extract global contextual features and local detail features, respectively. The key semantic feature selection module adaptively filtered critical features based on the channel state signal-to-noise ratio (SNR). The joint source-channel encoding-decoding module transmitted only essential information to minimize redundancy. Experimental results on the NWPU-RESISC45 dataset demonstrate that, compared with state-of-the-art methods, the proposed method achieves higher transmission accuracy and robustness under varying channel types and SNR conditions.
The popularization of cloud storage and the growing demand for data ownership transfer have made remote data integrity verification with ownership transfer a key technology for safeguarding cloud data security. However, existing schemes were generally plagued by problems such as heavy computational burden on data owners, reliance on untrusted third-party auditors, high communication overhead, and lack of effective non-repudiation mechanisms. To address these challenges, a remote data integrity verification scheme with ownership transfer based on trusted execution environment (TEE) was proposed. All critical computational tasks in the integrity verification and ownership transfer are securely offloaded to the TEE of the cloud service provider’s platform by our scheme. Formal security analysis was conducted based on a strong threat model where all entities except the isolation protection of TEE were untrusted, and a prototype system was developed using Intel SGX. Comparative experiments with various representative schemes were performed to evaluate its performance. The experimental results demonstrate that this scheme achieves superior computational efficiency in stages such as homomorphic verifiable tag generation compared to the benchmark schemes, with computational overhead reductions ranging from 10.1% to 86.4%. Additionally, it achieves zero communication overhead between cloud service providers and verifiers.
Aiming at the pain points in current network operation and maintenance, a four-layer practical framework was constructed, comprising autonomous network capabilities, full-process network visibility, network event monitoring, and scenario-based agent applications. At the core algorithm level, a “GBDT-Expert Hybrid” fault localization algorithm was proposed, which transformed expert experience rules into machine-readable features and fused them with the GBDT model, constructing a “first demarcation, then localization” tandem mechanism to effectively resolve the sample imbalance problem and achieve precise tracing of fault root causes. Meanwhile, a highly decoupled and orchestratable atomic capability foundation was constructed, and large language model (LLM) technology was introduced to realize intelligent interaction. Practical results demonstrate that the agents enable end-to-end automated cross-domain processing, shortening the average fault handling time by 14.8 minutes.
Network attack-defense platforms serve as a critical infrastructure supporting cybersecurity exercise and the improvement of practical defense capabilities. The empowering pathways and systematic construction methods of artificial intelligence (AI) technologies for such a platform were focused upon. To address the limitations of traditional platforms in terms of dynamic evolution, scenario diversity, and evaluation dimensions, key technologies including multi-agent systems, reinforcement learning, and large language models (LLM) were integrated. A layered and configurable platform architecture was proposed, which integrated intelligent attack simulation, adaptive defense decision-making, and automated evaluation feedback. Furthermore, a hierarchical agent training mechanism was designed, which was based on imitation learning and lifelong learning. High-fidelity attack chains were dynamically generated by this platform, cross-domain collaborative defense and multi-dimensional quantitative evaluation were realized, and the practicality and training precision of cyber exercises were effectively enhanced. Through the scenario-based configuration of the policy library, environment simulation, and evaluation system, diverse application requirements, including education, scientific research, and industrial exercises, were flexibly accommodated. Key challenges faced by the platform, including explainability, simulation fidelity, data privacy, and ethical norms, were further analyzed. Future research directions were also envisioned from the perspectives of explainable artificial intelligence, digital twins, privacy computing, and standardized collaboration. Thereby, theoretical support and practical references were provided for the construction of a next-generation, adaptive, and sustainably evolving cyber defense system.
With the rapid progress of artificial intelligence and high-performance computing, graphics processing unit(GPU) has become the core resource of modern compute centers. Contemporary deployments typically integrate GPU from multiple vendors with diverse architectures. However, the traditional solutions based on hardware partitioning or driver binding are usually tied to a single vendor and thus fail to provide efficient sharing and isolation in such heterogeneous environments. The SandGPU,a heterogeneous GPU resource pooling solution with an API sandbox mechanism was presented. At the resource-abstraction layer, SandGPU built a unified three-dimensional virtual GPU model and introduced vendor-level normalization coefficients to equivalently characterize the compute capability, memory capacity, and bandwidth of different GPU architectures. At the execution layer, SandGPU employed an API-level sandbox with a memory-quota gate, bandwidth token bucket, and compute throttling gate to enforce quota-based rate limiting and suppress cross-task interference. The experiments on a self-developed discrete-time simulation platform show that, under comparable average GPU utilization and task completion ratio within deadlines, SandGPU reduces the P95 value of the interference ratio by an average of 22.4%, significantly tightening the interference tail.
With the acceleration of global digitalization, the coverage gap of terrestrial cellular networks across maritime, aviation, and remote areas has become increasingly prominent. To address this issue, an in-depth study of voice over Wi-Fi (VoWi-Fi) and satellite converged communication technologies was conducted. Firstly, the limitations of traditional voice over IP (VoIP) technology, 4G/5G voice services (voice over LTE/voice over New Radio, VoLTE/VoNR), and their integration schemes with satellite technology were analyzed. On this basis, a deep convergence solution integrating VoWi-Fi with the “space-ground network” satellite model was proposed based on the 3rd Generation Partnership Project (3GPP) S2b interface architecture, and a cross-domain service tracing mechanism was designed to meet communication security requirements. The results of laboratory system tests and validations in typical scenarios such as cruise ships and aviation demonstrate that the proposed solution achieves high voice quality, largely comparable to the VoLTE/VoNR experience, while requiring only approximately 300-400 kbit/s of satellite backhaul bandwidth per voice call. Effective global coverage for voice, video, and SMS services is achieved, along with ensured user experience and communication security.The study provides a practical and feasible technical pathway towards achieving low-cost ubiquitous communication coverage.
In mobile communications, in order to solve the problems of bad speech signal decoding quality under poor transmission conditions, an innovative error concealment technology based on soft-decision decoding algorithm was proposed. Different from the traditional hard-decision decoding and standard error concealment technology, log likelihood ratios reflecting the probabilities of errors per bit, were expected in the soft-decision decoding algorithm and used to calculate the posterior probability. The signal was reconstructed after parameter estimation. Simulation results show that in the adaptive multi-rate-wideband (AMR-WB) decoders, compared to traditional hard-decision decoding algorithms and standard error concealment technologies, with using the proposed soft-decision decoding algorithm, the mean opinion score (MOS) and gain can be significantly improved, which further indicates an improvement in speech quality. This method can also be easily replicated in other decoders.
To address the problem of task allocation in scenarios with dynamic task introduction and dynamically changing rewards, a distributed multi-agent task allocation algorithm named I_GRAPE was proposed based on anonymous hedonic games. The I_GRAPE algorithm was designed by integrating a Q-learning-based multi-objective weight adaptation strategy with a log-linear learning method. Agents assigned to the same task were regarded as a coalition, thereby transforming the task allocation problem into a coalition formation game, with the objective of maximizing system utility while minimizing the moving distance of agents and the time cost required to complete tasks. Experimental results demonstrate that the I_GRAPE algorithm achieves satisfactory performance in multi-agent task allocation, with low costs in terms of both moving distance and task completion time. Compared with classical task allocation algorithms, the comprehensive performance of I_GRAPE is improved by 5.85% to 19.14%. The superiority and stability of the I_GRAPE algorithm in practical applications are preliminarily validated.
The integrated sensing and communication based Internet of vehicles (ISAC-IoV) system that utilizes ISAC beams to achieve vehicle target sensing and information communication, has received widespread attention in recent years, because of its significant improvement in system spectrum efficiency. However, the limited wireless resources in ISAC-IoV systems make it difficult to simultaneously satisfy the quality of service (QoS) for both sensing and communication. Furthermore, a unified metric for evaluating system performance is lacking. To address these challenges, firstly, the concept of value of service (VoS) was introduced into the ISAC-IoV system. The VoS for sensing and communication was defined using mutual information and transmission rate, respectively, replacing the traditional QoS. Secondly, a wireless resource allocation scheme was proposed to maximize the weighted VoS of the ISAC-IoV system by jointly optimizing the base station (BS)-vehicle association and BS channel allocation. Since the optimization problem was an integer non-linear programming (INLP) problem and NP-hard, it was decoupled into two sub-problems: BS-vehicle association and sub-channel allocation. A dynamic preference revision mechanism was also introduced to improve traditional matching theory and the sub-problems were solved separately. Finally, the optimal solution to the original problem was achieved through cyclic iterations. Simulation results show that in the ISAC-IoV system, the proposed algorithm significantly improves the system VoS compared with the benchmark algorithms. Specifically, enhancements of approximately 25% and 40% are achieved at low and high vehicle densities, respectively. This verifies the superior performance of the proposed algorithm.
The rapid development of new energy urgently requires more user-side resources to participate in demand response to ensure the balance of supply and demand in the power system. Blockchain-empowered demand response can effectively reduce trust costs among market entities with various types and different electricity consumption habits. However, how to improve blockchain consensus throughput and reduce consensus delay by jointly optimizing security resource management strategies such as the block quantity selection, main node election, and sub-channel allocation, so as to satisfy the system’s demand for efficient transaction processing and low-delay interaction is a key issue that needs to be addressed urgently. Based on this, a multi-timescale intelligent security resource management method for blockchain empowered demand response was proposed. Firstly, a blockchain-empowered secure demand response framework was constructed to facilitate information exchange among grid companies, demand response aggregators, and electricity users. Secondly, a joint optimization problem was formulated with the objective of maximizing the weighted sum of consensus throughput and the reciprocal of block consensus delay. Finally, at a large timescale, the main node election and sub-channel allocation were jointly optimized based on reputation matching, small timescale, the optimization of block quantity selection was based on deep Q-learning network. Simulation results show that the proposed algorithm performs well in terms of consensus throughput, block consensus delay.
To meet the quality of service (QoS) requirements of emerging intelligent services such as large language model inference and multi-agent collaboration, constructing an efficient resource scheduling mechanism has become a core issue in AI computing networks. For the resource scheduling scenario under QoS multi-objective optimization, a scheduling framework that integrates differentiated intelligent service modeling, heterogeneous computing-network resource awareness, and high-throughput remote direct memory access (RDMA) transmission optimization was proposed. Within this framework, a reinforcement learning (RL)-based intelligent scheduling algorithm was developed to adaptively generate scheduling policies by considering service priorities, computing-network resource loads, and data transmission characteristics, thereby improving QoS. Finally, a simulation platform was constructed using real-world network measurement data and actual topology parameters from an AI computing center to verify the performance of the proposed algorithm. Experimental results demonstrate that the proposed method achieves significant improvements over existing schemes in key QoS metrics such as task completion rate, average latency, and resource utilization.
Aiming at the stringent delay and energy consumption requirements of computation-intensive tasks in the Space-Air-Ground Integrated Network (SAGIN), this paper explores how to rigorously map the complex heterogeneous physical characteristics of SAGIN into a Markov Decision Process (MDP) and introduces a multi-head attention mechanism to optimize the Multi-Agent Deep Deterministic Policy Gradient (AM-MADDPG) algorithm. Considering the curse of dimensionality in highly dynamic environments, a Centralized Training with Decentralized Execution (CTDE) architecture is introduced. Each edge agent utilizes multi-head attention during the feature extraction phase to accurately assess collision risks and achieves coordination with ultra-low delay. Simulation results show that the proposed mechanism effectively alleviates task delay constraints. Compared with mainstream benchmarks like MAPPO and distributed Graph Reinforcement Learning based on local neighborhood aggregation (DGN), the proposed algorithm demonstrates superior cost reduction under the premise of equivalent signaling overhead, enhancing system resilience in extreme heavy-load scenarios.
As critical infrastructure，IP networks are facing an urgent need for automation and intelligent transformation in change management and security protection. An intelligent agent O&M platform is developed in this paper. Addressing the issues of IP network change monitoring and router APT security defense, two high-value L4 scenarios of the autonomous network are selected: configuration change and network security. Two major intelligent agents are focused, where configuration generation, configuration verification, configuration simulation, and service comparison is built by the change agent, based on digital twins. Security configuration verification, exposure surface management, and system intrusion detection functions are constructed by the security agent. Through pilot tests on real networks, review time is compressed to 30 minutes and efficiency is improved by 90%. Weak configuration exposure periods are reduced to the daily level with 90% efficiency improvement. Post-event auditing of device intrusion threats can be achieved by minute-level awareness. The O&M risks and security incident rates can be effectively reduced, while network stability and security protection levels can be significantly enhanced.
Radio frequency (RF) fingerprint clustering is a core technology for the identification of wireless communication devices. However, the joint clustering of multi-domain signals collected across different receivers has not been effectively explored. To address this issue, a pseudo-label and domain adaptation network co-driven RF fingerprint clustering method was proposed. First, Simsiam (simple siamese, Simsiam) network was employed to mine features from unlabeled source-domain RF fingerprint data, and the K-means algorithm was used to generate pseudo-labels. Then, a domain adaptation network was constructed, which took the source-domain data with pseudo-labels and the unlabeled target-domain data as input to enable the model to learn domain-invariant features. Finally, the K-means algorithm was applied to cluster the domain-invariant features. An experimental dataset was constructed using RF signals collected from 8 USRP devices in a cross-receiver scenario. The results show that the proposed method achieves 99.72%, 0.9918, and 0.9936 in terms of recognition accuracy, normalized mutual information (NMI), and adjusted Rand index (ARI), respectively, which are 47.72%–49.72%, 0.33–0.49, and 0.57–0.64 higher than those of the three comparison methods.
As the Open Radio Access Network (O-RAN) for Sixth Generation mobile communication technology (6G) evolves toward intelligent autonomy, the "black-box" nature of highly relied-upon Artificial Intelligence (AI) models conflicts with the requirement for transparent and trustworthy decision-making. The deep integration of eXplainable AI (XAI) is essential to demystify this black-box, achieve transparent decision-making, and construct a trustworthy autonomous network. This paper provides a comprehensive review of XAI research in 6G O-RAN and proposes a multi-layer integrated deployment framework and roadmap: we expound on the O-RAN architecture and XAI deployment logic; analyze implementation paths and typical applications across radio resource management, network slicing, network security enhancement, intelligent collaboration, and zero-touch operation; discuss technical challenges, including the trade-off between explanation computational overhead and real-time requirements, the lack of unified evaluation standards, interface-induced coordination constraints, and data privacy risks; and explore future trends, such as lightweight XAI algorithms, neuro-symbolic causal reasoning models, closed-loop autonomous mechanisms, and interdisciplinary integration, offering insights for constructing intelligent, transparent, and secure 6G trustworthy autonomous networks.
The development of space–air–ground networks is critical for 6G communications. An ultra-wide-angle scanning full-polarization phased array antenna with corresponding beam synthesis method was proposed for application in complex satellite–terrestrial interconnection scenarios. The full-polarization element contained four independent channels with different polarizations and beam ranges. The flexibility of element beamforming in beam direction and polarization was realized by adjusting the feeding power weight and phase difference of the four channels. The full-polarization ultra-wide-angle scanning was realized by applying the element in phased array. A four-element linear array prototype was designed, fabricated and experimentally verified in the band from 2.35 to 2.45 GHz. The isolation between each port were better than 15 dB. The beam scanning of θ = ±90º for both linear polarization and circular polarization were achieved with low scan loss better than 3.9 dB. The excellent multi-polarization characteristics and low-loss ultra-wide-angle scanning capability can effectively enhance the quality and stability of satellite–terrestrial interconnections, thereby further promoting the development of space–air–ground integrated networks.
With the wide application of new energy power generation technologies, the types of power quality disturbances (PQD) in power grids are becoming increasingly complex, and the phenomenon of multiple disturbances superimposing frequently occurs. Traditional classification methods, which rely on complete training samples, are difficult to effectively identify unseen compound disturbance types. Therefore, a power quality disturbance classification method based on time-frequency feature fusion and confidence optimization (TFFF-CO) is proposed. Firstly, the PQD signal is subjected to fast Fourier transform (FFT) to obtain spectral information, and the time-domain dynamic features are extracted using the attention-enhanced gated recurrent unit (AGRU-ATT), while the frequency-domain local features are captured by the lightweight multi-scale convolutional neural network (MS-CNN-Lite). Secondly, the time-frequency features are fused through an adaptive weight allocation strategy to enhance the integrity and specificity of feature representation. Then, in the multi-label learning framework, category discrimination labels are introduced to achieve the preliminary division of single and multiple disturbances and predict the confidence of each disturbance label. Finally, a dynamic confidence optimization factor is designed to improve the confidence discrimination ability of unknown multiple disturbances without affecting the recognition accuracy of known disturbances. Simulation experiments show that when this method is trained only with single and double disturbance samples, the recognition accuracy for triple and quadruple disturbances reaches 99.42% and 96.38%, respectively.
