To address the challenges of error accumulation and long-horizon divergence in multi-step network forecasting for 5 G core and backbone transport networks, this paper proposes a method based on Future-Guided Learning (FGL) and Multi-Teacher Distillation. The proposed approach introduces a causally consistent framework that leverages futurewindow teachers and a history-window student model. More specifically, multiple heterogeneous teacher models generate predictive distributions based on short-horizon windows that are closer to the forecasting targets. Then time-shifted probabilistic distillation transfers the resulting future soft labels to the student model, which is trained solely on historical data. This method significantly improves the accuracy and stability of long-horizon predictions. It does not incorporate future data during inference. For the student model, we adopt iTransformer-Lite, a lightweight multivariate time-series architecture, which reduces model size while enhancing the efficiency of multivariate interaction modeling. Experiments on the real-world CESNET-TimeSeries24 dataset demonstrate that the proposed method achieves lower RMSE and higher R-squared values for long-horizon forecasting, while maintaining a lightweight inference process. This approach provides reliable forward-looking insights, supporting proactive operations, traffic engineering, and slice resource orchestration in 5G and beyond networks.
Affine frequency division multiplexing (AFDM) is a promising multicarrier waveform for high-mobility scenarios due to its robustness against doubly-dispersive channels. However, accurate channel estimation remains challenging because delay and Doppler shifts are coupled in the discrete affine Fourier transform (DAFT) domain. In this paper, unlike conventional AFDM channel estimation schemes that rely on a relatively large chirp parameter to maintain path separability in doubly-dispersive channels, we develop a CP-based channel estimation scheme without this requirement. This makes the proposed method more suitable for channels with large delay spreads, where conventional designs would otherwise suffer from high pilot overhead. Specifically, the cross-correlation between the cyclic prefix (CP) and the corresponding symbol tail is computed over multiple symbols and averaged to estimate the Doppler shifts. The path delays are then identified using pilots in the DAFT domain, based on which an equivalent channel matrix is constructed for DAFT-domain equalization. Crucially, we show that under random signaling, the phase of the CP cross-correlation depends only on the Doppler shifts and is invariant to multipath interference. Simulation results demonstrate that the proposed method achieves a Doppler normalized mean square error (NMSE) of $10^{-3}$ at an SNR of 17 dB. Moreover, the proposed CP-AFDM scheme demonstrates superior BER performance over conventional multicarrier waveforms.
We consider a hybrid near-field and far-field simultaneous wireless information and power transfer (SWIPT) system assisted by simultaneously transmitting and reflecting reconfigurable intelligent surfaces (STAR-RIS), where a fullduplex hybrid access point (HAP) serves multiple users via SWIPT. We formulate a sum-rate maximization problem, by jointly optimizing receive beamforming, energy beamforming and STAR-RIS coefficients whose solution is achieved via alternating optimization (AO) that combines penalty-based semidefinite relaxation (SDR), and an interior point method. Via simulations, the sum-rate performance of the proposed algorithmic solution is verified with the convergence analysis.
This work exploits the degrees of freedom offered by the null space of the time-varying cascaded channel between the base station and information users (IUs), as mediated by an intermediate reconfigurable intelligent surface (RIS), to transmit an additional energy signal to energy users (EUs) concurrently with the data transmission to the IUs. Specifically, an extended null space (ENS) based design is proposed for a RISassisted simultaneous wireless information and power transfer (SWIPT) system, enabling simultaneous information and energy transfer. The proposed framework formulates a multi-objective optimization problem aiming to jointly maximize the harvested power of the EUs and the data rates of the IUs. To solve this problem, a deep deterministic policy gradient (DDPG) algorithm is developed and implemented. Simulation results demonstrate that the DDPG-based ENS design consistently outperforms state-of-the-art SWIPT schemes by improving the IUs' sum-rate and the total equivalent sum-rate (ESR) of the entire system, while maintaining competitive performance for the wireless power transfer.
Reconfigurable intelligent surfaces (RIS) are increasingly recognized as an enabling component of sixthgeneration (6 G) wireless systems, as they allow controllable modification of the propagation environment to improve both sum-rate and energy efficiency. This work investigates a UAVmounted simultaneous transmitting and reflecting RIS (STARRIS) network, where an energy-harvesting-enabled aerial platform functions as a reconfigurable relay to support users located on both the reflection and transmission sides. A joint nonconvex optimization problem is formulated to optimize the UAV trajectory, base station beamforming, STAR-RIS reflection and transmission coefficients, and the energy harvesting powersplitting ratio. The objective is to maximize a weighted sum of energy efficiency and harvested energy while satisfying users' quality-of-service (QoS) requirements under practical mobility and transmit power constraints. To address this challenging control problem, deterministic policy-gradient-based deep reinforcement learning algorithms, namely deep deterministic policy gradient (DDPG) and twin delayed DDPG (TD3), are employed within a continuous-control environment that incorporates realistic Rician fading channels. Simulation results demonstrate that TD3 achieves more stable convergence behavior and superior energy efficiency compared to the baseline DDPG approach while effectively optimizing the weighted objective of energy efficiency and harvested energy.
Unlike conventional wireless systems tolerating higher transmission delays, emerging delay-sensitive and mission-critical applications demand ultra-low latency and high reliability. In such scenarios, impairment like misalignment errors (ME) and transceiver hardware impairments (THI) critically affect system performance and compromise quality of service (QoS). Thus, this work analyzes the effective rate of a THz-based NOMA system considering ME and THI effects. Closed-form probability density function (PDF) and cumulative density function (CDF) are derived using Fox's Hfunction to characterize the channel behavior under various THz fading conditions accurately. Interestingly, although the PDF of the instantaneous signal-to-noise ratio appears relatively simple in form, its application in analyzing wireless system performance metrics becomes mathematically intricate. Hence, for the effective rate analysis, we adopt the same methodological framework as presented in [1]. Both fixed and adaptive antennaarray configuration are evaluated, demonstrating that adaptive optimization significantly enhances the effective rate. Finally, the presented analytical results are thoroughly validated through extensive simulation, where their accuracy is confirmed by close assessment with the corresponding Monte-Carlo simulation results.
Wireless communication systems offer strong potential for accurate localization, and deep-learning-based fingerprinting has shown good adaptability in complex propagation environments. However, existing fingerprints are limited to static location features, and current neural network designs do not fully utilize their structural properties. To improve this, we introduce a dynamic triple-beam fingerprint (TBF) for massive multipleinput multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) systems and develop a localization and orientation awareness network (LOA-Net) designed to align with its sparse structural characteristics for joint localization and orientation estimation. We analyze the advantages of TBF for localization and orientation awareness through channel modeling. Localization is formulated as a regression task and enhanced with a masking mechanism, while orientation estimation is modeled as a multi-class classification problem using the estimated coordinates as prior information. Simulations in standard 3GPP scenarios demonstrate the high localization accuracy and the potential for effective orientation awareness.
The increasing complexity of dense 6G RAN environments necessitates a paradigm shift from conventional linklevel optimization toward system-level control mechanisms that ensure intelligent coordination of radio resources, mobility, and device management across multiple coexisting RANs. Traditional optimization techniques, including advanced beamforming, modulation schemes, and interference mitigation, operate within isolated link boundaries, lacking the global awareness necessary for dense deployments. In contrast, system-wide AI-driven control frameworks enable holistic optimization, dynamically adjusting network parameters to adapt to fluctuating demand, user mobility patterns, and energy efficiency constraints. This paper proposes an AI-native system-level RAN control architecture, which redistributes core network functionalities to the edge and introduces a flexible deployment model: as part of C-RAN, as an independent edge core, or within the centralized network infrastructure. The proposed approach enhances power and frequency management, optimizes paging and mobility execution at the edge, and reduces core signaling overhead by executing localized resource coordination and interference control. AI plays a fundamental role in this transition, enabling real-time adaptive decision-making tailored to localized network conditions. By shifting from isolated, link-level improvements to a coordinated, AI-driven system-level optimization paradigm, the proposed framework enables efficient, scalable, and resilient dense RAN deployments. Future work will focus on prototype development and real-world validation, assessing the impact of AI-based RAN management and local mobility execution while ensuring seamless integration with existing 6G network architectures.
This survey examines the privacy and security challenges of deploying generative AI at the network edge. Unlike cloud-based systems, edge GenAI is susceptible to hardware-level exploits and training-data leakage due to physical accessibility and heterogeneous security boundaries. We present a comprehensive threat taxonomy and systematically evaluate documented attacks-such as power analysis and cache side-channel techniques-across mobile and industrial IoT platforms. We also analyze the efficacy of Trusted Execution Environments (TEEs) and selective-protection schemes, highlighting a 3-21% CPU overhead associated with hardware isolation. The paper concludes by outlining open research directions for securing multimodal models and scaling enclaves for billion-parameter deployments.
In this paper, we investigate energy-efficient beamforming and power control for network-assisted full-duplexenabled cellular-connected unmanned aerial vehicle (UAV) communications supporting heterogeneous uplink/downlink (UL/DL) services; E-UAVs upload high-rate payloads in the UL, while UUAVs receive low-latency control signals in the DL. We consider a lightweight UAV antenna architecture with a small number of phase shifters (PSs) for antenna on/off control and low-resolution analog-to-digital/digital-to-analog converters (ADCs/DACs). Then, we formulate an EE maximization problem by jointly optimizing PS deactivation and the E-UAV transmit power, subject to heterogeneous UL/DL QoS requirements. To reduce channelestimation complexity and signaling overhead (SO), we propose a distributed heterogeneous graph neural network. Simulation results demonstrate that the proposed scheme achieves higher EE with lower SO and complexity than existing schemes.
Radio-frequency fingerprinting identification (RFFI) can strengthen wireless security, but studies in 5G New Radio (NR) are scarce, and performance often degrades under channel variation. We propose a channelrobust RFFI framework for 5 G NR. First, we introduce a Channel-Independent Differential Operator (CIDO) that derives channel-invariant device features from NR demodulation reference signals (DMRS), suppressing channel distortion while preserving device traits. Second, to accommodate NR's flexible resource mapping and exploit frequency-domain dependence, we design the Frequency-domain Autocorrelation Network (FANet), which handles variable-length frequency inputs and employs frequency-domain attention for improved discrimination. Finally, using data from five commercial 5G NR devices collected across diverse environments and times, our method attains consistently high accuracy and markedly better cross-channel generalization than the baselines. These results indicate a practical path toward reliable RFF-based device identification in dynamic $\mathbf{5 G}$ NR settings.
Integrated Sensing and Communication (ISAC) has emerged as a key enabling technology for next-generation wireless networks. However, conventional ISAC systems based on orthogonal frequency division multiplexing (OFDM) suffer performance degradation and limited sensing resolution under multipath propagation. To overcome these limitations, this paper proposes an ISAC framework built on orthogonal chirp division multiplexing (OCDM), whose waveform structure provides enhanced robustness against multi-path interference. Furthermore, index modulation (IM) is incorporated into the OCDM-ISAC design to implicitly convey additional information bits through activated subcarrier indices, thereby improving spectral efficiency. At the receiver, a joint equalization and lowcomplexity log-likelihood ratio (LLR)-based detection scheme is employed for communication symbols, whereas a highaccuracy cyclic cross-correlation (CCC)-based algorithm is adopted for target parameter estimation in radar sensing. Simulations demonstrate that the proposed IM-OCDM-ISAC system outperforms the conventional OFDM-ISAC, providing more reliable communication performance in multipath channels and achieving higher sensing resolution.
Efficient processing and transmission of multisource visual data are critical for unmanned aerial vehicle (UAV)-based applications in complex environments. However, transmitting high-volume visible (RGB) and infrared (IR) images over bandwidth-limited wireless channels remains challenging. This paper proposes a Task-Oriented Adaptive Semantic Communication framework (TAMF-SC) with Multi-Modal Fusion to address the challenges of high-volume data transmission and low signal-to-noise ratio (SNR) conditions in UAV RGB-IR systems. The framework adopts an end-to-end architecture that performs cross-modal fusion and semantic encoding of RGB and IR features, transmitting semantic features for downstream tasks and high-quality fused image reconstruction. An SNR adaptive module dynamically adjusts the per-channel feature gains according to the estimated channel SNR, emphasizing critical semantic features while suppressing noisy ones, thereby enhancing robustness under varying channel conditions. At the receiver, the received semantic features are first decoded and then utilized for both fused image reconstruction and object tracking. Simulation results show that the proposed TAMF-SC framework outperforms representative baselines, achieving improved object tracking accuracy, measured by Mean Success Rate (MSR) and Mean Precision Rate (MPR), as well as high-fidelity fused image reconstruction under Additive White Gaussian Noise (AWGN) and Rayleigh fading channel scenarios.
To address the challenges of limited sensing range, high computational overhead, and insufficient real-time performance in path planning for automated guided vehicles (AGV) systems in dynamic industrial scenarios, this paper proposes a perception-driven conflict-based search (P-CBS) algorithm. This method constructs a sensing cost mapping model within an integrated sensing, communication, and control (ISCC) framework, which integrates sensing, communication, and control functions into wireless networks to achieve reciprocal enhancement. This allows the system to adaptively balance sensing update frequency against communication load during path control. Specifically, it converts visual detection information such as position, speed, and confidence into a confidence-weighted spatiotemporal Gaussian cost field. Soft constraints are introduced in the $\mathrm{A}^{*}$ search algorithm to achieve continuous obstacle avoidance. Furthermore, an event-triggered replanning mechanism based on cost growth rate prediction is designed, initiating global planning only when high risks accumulate, reducing computational load. Finally, a communication-adaptive threshold adjustment strategy is introduced to maintain system stability and consistency in bandwidth-constrained environments. Simulation results show that compared with traditional CBS and global dynamic CBS algorithms, the proposed algorithm can reduce the frequency of global replanning by at least 83 % and improve system throughput by at least 12.5 %, verifying its efficiency and robustness in dynamic industrial scenarios.
Semantic communication focuses on the transmission of semantic information and aims to improve communication efficiency. The combination of semantic communication and Hybrid Automatic Repeat reQuest (HARQ) schemes, which play a crucial role in ensuring reliability, is thus logically reasonable. However, traditional HARQ schemes use cyclic redundancy check (CRC) for error detection, which may lead to significant overhead since the CRC focuses on the bit-level errors and adds redundant bits, while semantic communication mainly focuses on the reliability of semantic information. To address this mismatch, a novel HARQ scheme is designed for semantic communication systems. Specifically, we propose a Semantic Quality Discriminator (SQD) network to assess the semantic quality score of the reconstructed information. Critically, a learnable re-transmission threshold is jointly optimized with the SQD network to enable re-transmission decision-making. The receiver combines the packets using the chase combining (CC) scheme to achieve a diversity gain. Experimental results demonstrate that our approach significantly increases communication quality while reducing communication overhead by optimizing the threshold for the desired performance trade-off.
Space-Air-Ground Integrated Networks (SAGINs) require sensing infrastructures covering sub-MHz to microwave frequencies. While Rydberg sensors offer ultra-wideband capabilities, a unified framework for concurrent detection remains unexplored. This paper proposes a unified atomic receiver simultaneously recovering microwave waveforms via LO-assisted mixing and sub-MHz fields via bias-assisted Stark modulation. We develop a linearized equivalent baseband model that transforms stiff time-domain dynamics into a rapid frequency-domain solution. Our analysis reveals a fundamental trade-off: the strong LO field required for microwave mixing induces Autler-Townes broadening, suppressing the spectral slope essential for lowfrequency detection. Consequently, we formulate a systematic strategy to reconfigure the receiver for microwave-priority, low-frequency-priority, or balanced sensing. By unifying disparate modalities within a single quantum aperture, this work offers a vital solution to reduce hardware complexity for size-constrained integrated sensing and communications platforms.
Low Earth Orbit (LEO) multibeam satellites with beam-hopping (BH) enable flexible resource allocation, but spatio-temporally non-uniform traffic and heterogeneous service priorities pose significant challenges to beam coverage and scheduling design. This paper proposes a user-priority-driven BH scheme that embeds service priorities into both dynamic beam footprint construction and time-slot allocation, with the objective of maximizing the weighted number of satisfied users. A two-stage heuristic algorithm is developed, where beam footprints are first generated using a minimum disk-coverbased method, followed by a priority-aware BH scheduling strategy that jointly considers user priority and residual traffic demand. Simulation results demonstrate that the proposed scheme significantly improves high-priority user satisfaction and achieves higher overall weighted satisfaction compared with fixed coverage and priority-agnostic schemes.
Recent advances in wireless networks have spurred interest in using Wi-Fi signals for human-centric mobile sensing tasks, such as gesture recognition and activity detection. To address the scarcity of labeled data in these applications, researchers increasingly turn to self-supervised learning (SSL) techniques for representation learning. In parallel, deep generative models, e.g., generative adversarial networks (GANs) and diffusion models, have been used to expand wireless sensing datasets via synthetic data generation. However, these efforts often treat generative models as black-box data expanders, with limited attention to semantic context or augmentation quality. In this paper, we propose a novel augmentation method termed the semantic augmentation via latent-space exploration (SALxO), which leverages the latent space of a variational auto-encoder (VAE) to generate semantically consistent views for contrastive learning. SALxO uses a latent chaining algorithm that traverses structured regions of the latent manifold, producing meaningful positive pairs without labels or handcrafted perturbations. Integrated into fixed SimCLR and AutoFi backbones, SALxO improves few-shot accuracy by up to 13 % and accelerates convergence, without requiring architecture changes. These results highlight SALxO's promise for scalable and label-efficient SSLenabled wireless sensing.
The reliable deployment of mobile Brain-Computer Interfaces (BCI) is fundamentally constrained by the perceptual nature of Electroencephalography (EEG) acquisition, where the sensed neural signals evolve continuously due to physiological, cognitive, and environmental factors, as well as the limited resources of wearable devices. While server-assisted adaptation can mitigate such perception-driven concept drift, it imposes significant communication and computational overhead. Existing solutions typically rely on rigid, periodic update schedules that fail to dynamically balance decoding accuracy against system costs. To address this, we propose a Drift-Aware Device-Server Framework driven by Reinforcement Learning (RL). In this architecture, an RL agent monitors distributional shifts in the perceptually acquired EEG signals and autonomously selects from a tiered set of maintenance actions, ranging from no-update to varying degrees of model retraining. By decoupling continuous sensing from expensive computation, the learned policy optimizes the longterm trade-off between inference reliability and resource consumption. Experimental results demonstrate that our framework outperforms static baselines, reducing computational load by $\mathbf{7 4 \%}$ and data transmission by 20 % while achieving superior decoding accuracy. This establishes a scalable, energy-efficient paradigm for sustaining robust BCI operation in resource-constrained wireless networks.
