Low-Earth-orbit (LEO) mega-constellations are expected to support ubiquitous connectivity, yet resilient access remains challenging under satellite mobility, adversarial jamming, and limited onboard resources. This paper investigates user association and scheduling in LEO networks under adversarial jamming, while accounting for link activation cost, finite radio frequency chains, and long-term queue stability requirements. We propose a decentralized online learning framework, termed Lyapunov-guided variance-aware multi-armed bandit (LV-MAB). Specifically, Lyapunov drift analysis converts the long-term queue stability requirement into a per-frame queue-weighted surrogate problem, which enables distributed user-side association and satellite-side scheduling. For user association, LV-MAB uses variance-aware upper confidence bound with link-level sliding windows and resets to track non-stationary rate statistics induced by mobility and dynamic jamming. For satellite-side scheduling, combinatorial Thompson sampling with posterior inflation is adopted to resolve contention among requesting users under radio frequency chain constraints. We establish dynamic regret bounds for LV-MAB and prove system queue stability. Extensive simulations show that LV-MAB achieves reliable throughput, backlog control, fairness improvement, and robustness under dynamic adversarial jamming.
Federated learning enables privacy-preserving model training in edge and communication networks with distributed and bandwidth-constrained clients. However, practical deployments often suffer from extremely limited labeled data and severe non-IID distributions, which hinder data-efficient and stable federated optimization. This paper proposes FSSG, a communication-efficient federated semi-supervised learning framework that combines client-side self-supervised representation learning with server-coordinated semi-supervised optimization. To improve robustness under network-induced heterogeneity, FSSG adopts a confidence-weighted aggregation mechanism based on prediction uncertainty evaluated on a small server-side labeled probe set. In addition, a dynamic residual coordination strategy is introduced to stabilize global model updates across communication rounds. Experimental results on multiple benchmark datasets demonstrate that FSSG consistently outperforms existing federated semi-supervised learning methods while maintaining moderate communication and computational overhead, making it suitable for data-efficient learning in edge networks.
The rise of cognitive electronic warfare has made synthetic aperture radar (SAR) systems increasingly vulnerable to active jamming, demanding intelligent perception and decisionmaking under uncertainty. This work formulates SAR jamming recognition as a cognitive sensing problem and proposes a Dynamic Feature-Fusion Network (DF2Net) tailored for Few-Shot Open-Set SAR Jamming Recognition (FSOS-JR). DF2Net employs a dual-stream attention extractor that jointly models global spatial-frequency patterns and local channel semantics, enabling robust representation learning from limited supervision. A dual-branch interaction module integrates a global-stream prototypical classifier for known-class identification with a localstream uncertainty estimator for rejecting unknown jamming types. Furthermore, an energy-based confidence metric fuses semantic and spatial cues into a unified scoring mechanism. Extensive evaluations on both the simulated JamSet and the real-world Sen1-Jam benchmarks demonstrate that DF2Net consistently achieves state-of-the-art performance across diverse jamming conditions, highlighting its strong generalization, openset reliability, and practical utility for real-world electromagnetic threat scenarios.
Neural channel decoders have attracted growing attention as promising routes for high-performance decoding of short block codes. A recent milestone is denoising diffusion error correction codes (DDECC), which fuses diffusion probabilistic models with a Transformer-based denoiser and outperforms earlier neural approaches. Yet its customized line-search update lacks formal convergence guarantees, and its convergence speed and decoding accuracy can still be improved. To overcome these limitations, we propose time-conditioned diffusion (TCD) decoder, a continuous-time neural decoder built upon stochastic differential equation (SDE) theory. TCD leverages high-order diffusion samplers to achieve theoretically-grounded, fast convergence. It further introduces (i) a time-conditioned Transformer-based denoiser with continuous embeddings, (ii) an initialization that maps received signals to valid diffusion states, and (iii) an early-stopping rule tailored for decoding. Moreover, we derive three analytically driven, low-complexity guidance schemes—likelihood-gradient, posterior-mean, and their hybrid—conditioned on the received signal to further enhance decoding accuracy. Simulations on BCH, LDPC, Polar codes show that guided TCD outperforms DDECC across nearly all settings while markedly reducing computational cost. Against other universal neural decoders, TCD achieves state-of-the-art performance, further closing the gap to maximum-likelihood decoding.
Future 6G networks will enable pervasive Generative Artificial Intelligence (GenAI) services. Migrating Large Language Model (LLM) inference to the network edge is essential to meet the low-latency and privacy demands of these services. However, LLM inference offloading faces a mismatch between resource-limited edge infrastructure and volatile computational demands. Conventional offloading schemes for static bit-level tasks cannot sufficiently handle the prefill/decoding phase difference and the uncertain output length of LLM requests. To bridge this gap, we derive a stochastic workload model based on decoder-only Transformer mechanics and profiled input/output token statistics, characterizing expected workload and task-dependent fluctuations. We embed these statistics into edge-cloud queue dynamics and formulate the massive-scale collaborative inference problem as a Mean Field Game (MFG). We approximate the high-concurrency queue dynamics using Stochastic Differential Equations (SDEs) and develop a Mean Field State Deep Q-Network (MFSDQN) algorithm. MFSDQN enables edge agents to learn decentralized offloading-ratio decisions under Quality of Service (QoS) constraints using local observations and neigh-borassisted mean-field states, avoiding global edge-state exchange. Extensive simulations with up to 500 edge agents demonstrate stable convergence and scalability. MFSDQN reduces the average system cost and unfinished workload by 8.0% and 6.3%, respectively, relative to the best-performing baseline.
Deep learning-based cognitive radio systems achieve strong results on modulation classification and SNR estimation, yet most models remain task-specific and depend on substantial labeled data. Large language models (LLMs) offer a more general interface, but applying them to radio signals requires bridging the gap between continuous I/Q sequences and discrete tokens. We propose RadioLLM-V2, an LLM-based framework for joint modulation classification and SNR estimation in few-shot settings. Distinct from traditional numerical position vectors, the proposed framework introduces a Semantic Positional Prompting (SPP) mechanism. SPP converts patch indices and sample-level time windows into language-driven positional cues, making patch-level temporal information more explicit in the LLM-compatible representation. The framework encodes I/Q samples with a dual-domain tokenizer and augments tokens with these semantic cues. Global signal statistics are converted into a compact physical prompt for prototype retrieval, and a prototype-guided Q-Former compresses the signal–prototype context into query tokens for a parameter-efficiently adapted LLM backbone. Experiments on five public benchmarks under the 100-shot protocol show that the proposed method achieves the best overall performance among the compared methods. On RML2016.10a, it improves classification accuracy from 58.33% to 59.29% and reduces SNR estimation RMSE from 3.12 to 2.74 compared with the original RadioLLM. It also attains 57.17% accuracy on the challenging RML2018 dataset, outperforming representative CNN/Transformer baselines.
Artificial intelligence (AI) models deployed on low Earth orbit (LEO) satellites can facilitate on-orbit intelligent information processing, effectively reducing communication overhead between remote sensing satellites and ground stations. In this paper, we propose a novel large-small model collaboration scheme over LEO satellite networks to address the deployment challenge of large AI models on resource-constrained remote sensing satellites, which enables delay-efficient on-orbit collaborative processing and small AI model updating. In the proposed scheme, remote sensing satellites process collected sensing data using small AI models and generate data packets for collaborative large AI model processing. Specifically, the data packets consist of extracted feature data and residual mapping data, which are transmitted to computing satellites through inter-satellite links. In addition, the computing satellites generate model packets and transmit them to remote sensing satellites to update the small AI models. Furthermore, we formulate an average task processing delay minimization problem via jointly optimizing the task offloading and routing under processing accuracy constraints characterized by mean average precision. We design a multi-agent reinforcement learning (MARL)-based two-layer algorithm. Specifically, the outer layer adopts an online bisection searching algorithm to iteratively search for the optimal task allocation ratio, while the inner layer utilizes the offline trained MARL policy to make routing decisions. Simulation results based on real-world remote sensing tasks demonstrate that the proposed scheme can reduce the average processing delay by 8.42% compared to the other benchmarks.
The evolution of 6G networks from bit-oriented transmission to semantic communication renders communication systems vulnerable to emerging security threats targeting information interpretation, specifically semantic distortion. Consequently, identifying critical attack paths is essential for understanding and mitigating semantic attack propagation at the network level. Reinforcement learning (RL) has been increasingly adopted to identify critical paths in complex networks, yet existing solutions typically rely on agents tailored to specific environments characterized by discrete nodes and fixed semantic attributes. These discrete formulations constrain scalability in large-scale networks and hinder generalization across dynamic topologies. To alleviate these limitations, we propose Sem-DRL (Semantic-aware Deep Reinforcement Learning), a continuous RL framework featuring invariant observation and action spaces. Sem-DRL leverages Graph Neural Networks (GNNs) to extract permutation-invariant embeddings, which enables zero-shot generalization across unseen network topologies. By decoupling the action space from the network size, the proposed framework ensures scalability in large-scale networks. Furthermore, Sem-DRL utilizes PLMs as semantic judges to quantify distortion rewards within a continuous latent space. Extensive experiments demonstrate that the proposed framework achieves stable convergence in networks with up to 500 nodes and attains zero-shot transfer success rates of 76%–90% across previously unseen network topologies.
Deploying Large Language Models (LLMs) over the edge-cloud continuum faces severe stability challenges due to the conflict between stochastic network topology and complex workflow dependencies. Existing schedulers, relying either on computationally prohibitive Graph Neural Networks (GNNs) or topology-agnostic heuristics, fail to reconcile this tension. To bridge these gaps, we propose STEM, a service-level and topology-aware orchestration framework that formulates distributed LLM serving as a workflow-aware routing problem over a monitored service overlay, in which heterogeneous service instances act as specialized experts. At the core of STEM lies the STAR-PPO algorithm, utilizing a lightweight graph-free perception mechanism. By leveraging Squeeze-and-Excitation attention, it extracts critical bottleneck features from raw telemetry with linear complexity, bypassing the scalability limits of message-passing paradigms. To further achieve Pareto-efficient trade-offs, we develop a Dynamic Weight Adaptation (DWA) mechanism that autonomously recalibrates optimization preferences based on entropy-regularized metric drift. Extensive experiments on real-world datasets spanning 2,000 nodes demonstrate that our framework significantly outperforms state-of-the-art baselines. Specifically, STAR-PPO reduces network transmission costs by 96.8% and improves comprehensive inference efficiency by 24.4%, while sustaining robust zero-shot generalization across regions, with average latency within 1.09× of a target-domainre-trained reference under a strict cross-region protocol. Code and data are available at https://github.com/gymorsiback/STARPPO.
Covert communication is a secure communication method that enables the transmission of secure information with a low probability of detection. In cases of constrained ground communication, unmanned aerial vehicles (UAVs) serve as ideal relay nodes for covert communication. Owing to their high mobility and ease of deployment, UAVs can significantly improve covert signal transmission in scenarios where communication links are obstructed or communication signals are weak. In scenarios where UAVs act as relay nodes and employ transparent forwarding to assist in covert communication, we have established a specific optimization problem. This problem, which aims to maximize system throughput through the joint optimization of transmission power, UAV amplification gain, and the UAV’s three-dimensional (3D) hovering position, exhibits non-convexity. Leveraging the rigorous constraint-handling capability of geometric programming (GP) and the strategic exploration strength of deep reinforcement learning (DRL) in dynamic settings, this paper proposes a hybrid optimization framework, namely the GP-based proximal policy optimization (GPPO) algorithm. Simulation results demonstrate that, compared with benchmark schemes, our algorithm significantly enhances system throughput, ensuring the efficiency of covert information transmission. This hybrid approach offers a promising solution for UAV-assisted covert communication.
Vehicle-to-everything (V2X) techniques expand the capability boundaries of connected and autonomous vehicles (CAVs). However, deploying V2X-enabled end-to-end autonomous driving (E2E-AD) systems still faces a trade-off between sharing high-resolution perception features and V2X communication bandwidth constraints. Furthermore, complex E2E models impose heavy computational overhead and inference latency on CAVs. To address these challenges, we propose EdgePlanner, an agentic V2X collaborative framework for CAV trajectory planning. EdgePlanner decouples conventional vehicle-centric E2E-AD into role-specialized on-board and edge agents, assigning explicit decision roles within a closed-loop planning workflow. The on-board agent is dedicated to preliminary perception processing. It employs clustering-based feature compression and temporal difference transmission modules to minimize bandwidth consumption while preserving critical semantic information. Conversely, the edge agent serves as a global planner that integrates collaborative V2X information from multiple CAVs, roadside infrastructure, and map context. It adopts an Agent Query module to capture complex interactions and generates highly reliable trajectories for on-board agents with a conditioned Denoising Diffusion Implicit Model (DDIM) decoder. Extensive experiments demonstrate that EdgePlanner consistently generates high-quality trajectories for CAVs, while significantly reducing communication overhead.
Federated learning (FL) applications normally employ large deep learning (DL) models, resulting in excessive communication overhead in the deployment of FL over resource-constraint mobile edge networks. To achieve better scalability for DL-based FL, we capitalize on both the asymmetric nature of mobile networks and the distinct effects of partial transmissions on FL training for the global and local models. We propose Fed-DynAmal, an FL framework that decreases the number of parameters transmitted in the uplink (clients-to-server) while concurrently achieving better model performance. The underlying idea is that each selected client sends a partial DL model to the server by omitting several sub-blocks from the trained local model. Crucially, we drop the assumption that transmitted local models can still be used for inference, thereby allowing for greater model variability. At the server, we introduce amalgamation, a process to merge different partial local models into an inference-viable full model. Essentially, amalgamation is a bridge for performing aggregation at the sub-block level. Interestingly, as the key takeaway, communication efficiency versus model performance is not necessarily a trade-off in FL:Our extensive experiments show that Fed-DynAmal can effectively improve communication efficiency while still concurrently achieving higher accuracy and enhanced robustness.
Semantic communication, delving into the essence behind bits, holds promise for facilitating information exchange across languages and domains. Despite its promising performance, deploying semantic communication on resource-constrained Internet of Things (IoT) devices faces significant challenges, including limited computational power, low semantic fidelity, and poor environmental robustness. To address these issues, this paper proposes a robust semantic communication scheme with enhanced fidelity, called LRSC. Firstly, to tackle the challenge of limited computational resources, a global-detail semantic enhancement (G-DSE) network is developed as a semantic encoder to achieve accurate semantic extraction, along with a semantic decoder based on the recurrent self-attention (RSA) mechanism for high-precision semantic interpretation. Second, to tackle the problem of low semantic fidelity, a semantic fidelity enhancement (SFE) network is developed, which incorporates an adaptive personalized mechanism based on redundancy suppression and contribution-based weighting. Finally, to improve physical-layer robustness, a semantic-driven singular value decomposition (SVD) signal optimization algorithm is proposed, along with an edge-terminal asynchronous training and online deployment strategy. Simulation results demonstrate that the proposed scheme achieves robust and high-fidelity semantic extraction and interpretation. Compared with existing state-of-the-art lite methods, it improves the PSNR by 2.8 dB and semantic fidelity by 8 %.
This paper proposes a novel terrain-aware, two-layer, graphics processing unit (GPU)-accelerated framework for base station placement optimization. In the first layer, three-dimensional (3D) K-means clustering is employed for initial base station seeding using real geographical, population, and elevation data, while two-dimensional (2D) K-means is applied in scenarios without elevation information. The second layer utilizes the quantum-inspired Raute optimizer (QRO), an adaptive hybrid metaheuristic that refines the initial placements. QRO dynamically alternates between exploration and exploitation through quantum-governed role switching and incorporates a novel inter-algorithm “bartering” mechanism for solution exchange with other heuristics. To validate the proposed framework, a comprehensive benchmarking study is performed on both terrain-aware 3D and 2D configurations. Within each framework, QRO is rigorously compared against six established metaheuristics. The GPU-accelerated simulations adopt the 3GPP rural macro (RMa) model and assess population coverage, convergence, and runtime. Additional experiments confirm robustness to spectrum-sensing errors and linear scalability with massive user densities. Results show that QRO—especially with bartering—consistently achieves higher coverage with faster convergence than competing metaheuristics. Multi-objective evaluation further demonstrates tunable trade-offs between coverage maximization and interference mitigation for practical network planning.
Existing deep joint source-channel coding (JSCC) methods typically employ uniform resource allocation across spatial regions, which may lead to spatially imbalanced reconstruction quality under limited channel resources. To address this, this paper proposes AgentJSCC, a difficulty-aware transmission framework. Specifically, a vision-language model (VLM)-based transmission agent is introduced to predict patch-level reconstruction difficulty, generating a difficulty mask conditioned on the communication setting. Guided by this mask, a region-adaptive JSCC encoder dynamically allocates transmission rates, assigning higher resources to challenging regions before transmission. To further mitigate residual artifacts, we extend this framework to AgentJSCC+, which incorporates a two-step edge-guided denoising network where EdgeNet extracts multi-scale edge priors and DenoiseNet performs edge-guided RGB refinement. Experimental results on AWGN and Rayleigh fading channels demonstrate that AgentJSCC improves reconstruction fidelity over existing JSCC baselines. On DIV2K over the AWGN channel, AgentJSCC improves PSNR over the strongest baseline by 0.87 dB on average, while AgentJSCC+ further increases the gain to 1.77 dB PSNR and an 11.4% relative SSIM gain, demonstrating superior robustness across diverse SNR conditions and compression rates.
This paper proposes a jointly adaptive AGC and parallel code-phase acquisition framework for airborne SAR communications under dynamic signal conditions. Unlike conventional receivers that optimize AGC and acquisition separately, the proposed architecture couples gain control, acquisition state, and noise-aware threshold adaptation through a closed-loop mechanism. A modulation-aware cascaded AGC strategy maintains ADC operation and front-end linearity over wide input variations. An FFT/IFFT-based parallel acquisition scheme reduces synchronization latency, while a high-noise-region adaptive threshold improves detection performance under nonstationary interference. FPGA experiments demonstrate reduced false alarms and missed detections, with low-latency acquisition achieved near the receiver sensitivity limit.
Cloud-edge-end collaborative Artificial Intelligence (AI) computing requires schedulers that allocate heterogeneous resources for Directed Acyclic Graph (DAG)-structured workflows across network tiers. Cross-tier data transfers create ripple effects where a single placement decision propagates delays to downstream tasks, degrading end-to-end completion rates. This paper presents DREAM, a Dynamic Ripple-Effect-Aware Meta-scheduling scheme in which Critical Path Lookahead Scheduling (CPLS) performs bounded-depth trajectory planning with soft reservations for critical tasks, while Opportunity-Cost-Aware Placement (OCAP) evaluates non-critical tasks through a four-component cost covering immediate efficiency, ripple effect, load stability, and opportunity cost. Extensive simulations demonstrate that under the heavy load of 600 tasks, DREAM sustains a task completion rate of ∼66%, exceeding classical heuristics by over 10 percentage points. At the extreme load of 1000 tasks, the system utility score improves by 47% over HEFT. Robustness experiments verify competitive performance across multiple DAG topologies and estimation-noise levels.
Networked systems continuously generate heteroge-neous time series, including Key Performance Indicator (KPI) streams, logs, and spectrum measurements, whose interpretation is essential for automated monitoring, diagnosis, and control. Existing analysis approaches either rely heavily on labeled data specific to each deployment or fail to capture joint time-domain and frequency-domain characteristics that are common in communication signals. Motivated by these limitations for cognitive communications and network monitoring, we propose the Time-Frequency Multi-Task Network (TFMTNet), a self-supervised framework that provides a transferable representation module for the evaluated network telemetry tasks. TFMT-Net integrates a multi-scale time-frequency fusion backbone with three complementary pretraining objectives and provides lightweight task heads for anomaly detection, forecasting, and classification. Under a pretraining and adaptation protocol, the model is pretrained once and then adapted to target domains with limited labeled data. Empirical evaluation on ten anomaly detection datasets and additional public classification and forecasting datasets, including an Artificial Intelligence for IT Operations (AIOps) telemetry dataset, shows consistent cross-domain improvements under the evaluated settings. For cognitive communications, TFMTNet learns representations that can feed downstream reasoning and control modules, including Software-Defined Networking (SDN) decision making, spectrum management, and AIOps pipelines, thereby supporting the perception, reasoning, and control loop in networked systems.
Low probability of intercept (LPI) signal detection is essential to cognitive electronic warfare (CEW) systems, but it is challenged by the coexistence of impulsive noise and the low signal-to-noise ratios (SNRs) stemming from LPI signals’ high time-bandwidth product and low peak power, thereby degrading the performance of traditional detection. Although recent deep neural network (DNN)-based methods show improvements, they lack principled false alarm control under Neyman-Pearson criterion (NPC) and overlook the impulsive noise. Besides, they fail to incorporate characteristics of LPI signals, and provide limited interpretability. To fill these gaps, this paper proposes an NPC-guided DNN detection framework. The key idea is to utilize DNN-based posterior probabilities as test statistic, achieving principled false-alarm control under NPC. The network is trained to extract discriminative LPI signal features, and detection threshold is computed using noise-only data. Building on this framework, we develop an end-to-end LPI signal detection network (ETED-Net), which integrates multi-scale feature extraction with channel attention to capture weak LPI signal features under hybrid AWGN and impulsive noise. Extensive experiments demonstrate that, within proposed framework, ETED-Net achieves superior detection performance and generalization over traditional and DNN-based methods. Furthermore, we demonstrate ETED-Net’s interpretability by tracing decisions to underlying signal structure.
