Accurately classifying radar-communication signals remains a fundamental yet challenging task, as these waveforms often exhibit high variability and are easily distorted by noise, channel fading, and spectrum overlap. Conventional deep learning approaches, despite their remarkable progress, tend to capture either local spatial cues or long-range dependencies in isolation, thereby limiting their ability to achieve robust recognition in complex environments. To address these challenges, we propose SynerNet, a lightweight deep neural architecture that effectively integrates spatial diversity and cross-scale attention mechanisms. Specifically, SynerNet leverages the smoothed pseudo Wigner-Ville distribution to generate informative time-frequency representations, which retain high resolution while mitigating undesired cross-term interference. Building upon these representations, the network is enhanced by two key components: the Gated Spatial-Channel Unit module, which jointly models spatial and channel dependencies to selectively emphasize salient features while suppressing noise, and the Fractal Cross-Scale Attention module, which employs a hierarchical fractal-inspired attention scheme to preserve fine-grained details across multiple scales while ensuring global consistency. Simulation results on 12 waveform classes encompassing both radar and communication signals demonstrate that SynerNet achieves an average classification accuracy of 90.61% , with only 47K parameters and an inference latency of 0.552 ms, outperforming existing deep learning approaches. These results highlight the strong potential of SynerNet for real-world deployment in intelligent sensing and wireless communication systems under resource-constrained environments.

Domain shifts induced by variations in urban scenes and adverse weather conditions pose a major challenge for semantic segmentation models in traffic environments. While adversarial domain adaptation has shown effectiveness in reducing cross-domain discrepancies, most existing approaches operate at a single adaptation level and struggle to generalize across diverse degradation patterns. In this work, we propose DA2LF, a dual-level domain adaptation framework designed to improve segmentation robustness under adverse conditions. At the input level, DA2LF reduces low-level appearance gaps via spectral transfer and targeted data augmentation. At the feature level, class-conditional adversarial learning is employed to align semantic representations between source and target domains. To further enhance fine-grained structural consistency, we introduce a multi-resolution Laplacian pyramid that preserves boundary details and small objects. Extensive experiments on both synthetic and real-world benchmarks demonstrate the effectiveness of the proposed approach. DA2LF is evaluated on three standard domain adaptation benchmarks: Cityscapes to Foggy Cityscapes, Cityscapes to Rainy Cityscapes, Cityscapes to ACDC, and consistently outperforms state-of-the-art methods based on global and class-wise alignment.

The evolution from fifth-generation (5G) to sixth-generation (6G) networks is driving an unprecedented demand for advanced machine learning (ML) solutions. Deep learning has already demonstrated significant impact across mobile networking and communication systems, enabling intelligent services such as smart healthcare, smart grids, autonomous vehicles, aerial platforms, digital twins, and the metaverse. At the same time, the rapid proliferation of resource-constrained Internet-of-Things (IoT) devices has accelerated the adoption of tiny machine learning (TinyML) for efficient on-device intelligence, while large machine learning (LargeML) models continue to require substantial computational resources to support large-scale IoT services and ML-generated content. These trends highlight the need for a unified framework that integrates TinyML and LargeML to achieve seamless connectivity, scalable intelligence, and efficient resource management in future 6G systems. This survey provides a comprehensive review of recent advances enabling the integration of TinyML and LargeML in next-generation wireless networks. In particular, we (i) provide an overview of TinyML and LargeML, (ii) analyze the motivations and requirements for unifying these paradigms within the 6G context, (iii) examine efficient bidirectional integration approaches, (iv) review state-of-the-art solutions and their applicability to emerging 6G services, and (v) identify key challenges related to performance optimization, deployment feasibility, resource orchestration, and security. Finally, we outline promising research directions to guide the holistic integration of TinyML and LargeML for intelligent, scalable, and energy-efficient 6G networks and beyond.

The growing demand for spectrum efficiency in next-generation wireless networks, especially in vehicular environments, necessitates effective spectrum sensing (SS) techniques capable of managing the coexistence of technologies like fifth generation new radio (NR) and radar systems. This letter introduces SpecDiff, an innovative framework based on latent diffusion models for spectrogram segmentation, designed to identify and differentiate these coexisting signals in dynamic, noisy environments. SpecDiff leverages a generative diffusion model in a compact latent space, using an attention-based denoising process to enhance segmentation performance under low signal-to-noise ratios and complex channel conditions. The model achieves state-of-the-art performance, with a mean accuracy of 98.68% and mean intersection-over-union (IoU) of 96.30%, effectively identifying the occupied bandwidth in spectrograms. Furthermore, SpecDiff surpasses existing deep learning models in both accuracy and efficiency, offering a promising solution for spectrum sharing in future wireless networks.

The increasing demand for efficient spectrum utilization in next-generation wireless networks necessitates advanced sensing methods to accurately characterize complex spectral environments. To address this need, this paper introduces HKANet, a novel encoder-decoder architecture designed for spectrogram segmentation. Its core innovation is centered on a dual-component design: a morphology-aware feature extractor that discerns diverse signal shapes, and a statistical attention block that utilizes higher-order statistics for robust feature recalibration. Extensive simulations on a diverse synthetic dataset comprising fifth-generation new radio, long-term evolution, and radar signals with varying bandwidths and modulation types demonstrate that HKANet achieves state-of-the-art performance, outperforming prominent models with a mean pixel accuracy of 98.49%, a mean intersection-over-union of 97.10%, and a mean F1-score of 98.45%. With a lightweight design of only 4.76M parameters and a competitive inference latency of 22.1 ms, HKANet presents an effective solution for intelligent spectrum sensing in future wireless communication systems.

Federated Learning (FL) enables distributed clients to collaboratively train models without sharing raw data, making it well suited for privacy-sensitive Internet of Things (IoT) environments. However, under heterogeneous (non-IID) data distributions, classical FL often suffers from unstable convergence and poor personalization, as repeated global aggregation overwrites client-specific knowledge. Existing Personalized Federated Learning (PFL) methods mitigate this issue through architectural partitioning or alignment strategies, but typically treat all parameters uniformly, leading to catastrophic forgetting. In this paper, we propose FedIPSA, a novel PFL framework that addresses this limitation through fine-grained, parameter-level personalization. FedIPSA integrates two modules: Importance-Based Parameter Masking (IPM) and Selective Alignment Mechanism (SAM). IPM estimates parameter importance using Synaptic Intelligence to identify critical client-specific parameters, while SAM preserves their local knowledge and selectively aligns the remaining globally initialized parameters with the previous local model. Experiments on CIFAR-10/100, FEMNIST, and Shakespeare demonstrate that FedIPSA outperforms state-of-the-art methods by up to 5.09% in large-scale settings while ensuring superior training stability and convergence. Furthermore, it achieves a 38.73% reduction in per-round execution time compared to recent alignment-based frameworks.

Compact spectrum sensing (SS) is crucial for optimizing wireless network performance, particularly within dynamic 5G NR and LTE environments. TThe effectiveness of traditional SS methods in real-world scenarios is often limited by challenges such as low signal-to-noise ratio, multipath fading, and high computational complexity. Deep learning-based approaches have emerged as promising solutions by offering superior adaptability and performance. However, existing DL models often suffer from computational costs and excessive storage requirements, thus constraining deployment on resource-constrained devices. To address these challenges, this paper introduces a novel Resolution-Preserving Network (RPNet), integrated with a Convolutional Layer Reuse (CLRu) technique, to enhance SS efficiency while reducing model size significantly. RPNet is designed to maintain the resolution of spectrogram input throughout its architecture, preserving critical spatial details for segmentation. CLRu optimizes memory usage by reusing convolutional layers, effectively reducing the number of trainable parameters with very slightly compromising performance. Extensive simulations on a synthetic 5G-LTE dataset demonstrate that RPNet+CLRu achieves a greater than 9 & times; reduction in model size while maintaining 87.45% segmentation accuracy. In addition, the proposed model outperforms existing deep models in storage efficiency and computational feasibility, hence presenting it as a promising candidate for real-time SS in next-generation wireless networks.

In recent years, data generation has grown exponentially, and Big Data has emerged as a propelling force in the development of various machine learning advances and Internet of Things devices. In this regard, the analytical and learning tools that transport data from several sources to a central cloud for processing, training, and storage enable the realization of the potential of Big Data. Nevertheless, since the data may contain sensitive information like banking account information, government information, and personal information, these traditional approaches often raise serious privacy concerns. To overcome such challenges, Federated Learning (FL) has emerged as a sub-field of machine learning that focuses on scenarios where several entities (commonly termed as clients) work together to train a model while maintaining the decentralization of their data. Although significant research efforts have been dedicated to this area, a comprehensive review focusing on FL within the realm of Big Data services is still lacking. This paper, therefore, emphasizes the use of FL in handling Big Data and related services, which provides a comprehensive review of the potential of FL in Big Data acquisition, storage, Big Data analytics, and further privacy preservation. Subsequently, the potential of FL in Big Data applications, such as smart city, smart healthcare, smart transportation, smart grid, and social media are also explored. The paper also highlights various projects related to FL for Big Data and discusses the challenges associated with such implementations. These discussions provide a direction for further research, encouraging the development of plausible solutions.

Multi-hop routing optimization in IoT networks involves complex energy-latency trade-offs that challenge traditional algorithm design approaches. This letter presents a systematic framework for Large Language Model (LLM)-assisted algorithm synthesis for multi-hop routing in digital twin (DT)-enabled IoT systems. The proposed three-phase methodology enables LLMs to support algorithmic design by analyzing problem structures, synthesizing hybrid optimization strategies, and iteratively refining routing heuristics under energy, reliability, and latency constraints. The framework generates multiple specialized algorithms, including Advanced Smart Constraint (ASC) for adaptive construction, Energy-Focused Iterative (EFI) refinement, and problem-specific large neighborhood search operators. Extensive experiments on synthetic IoT networks with K=5 -50 devices demonstrate consistent energy reductions of approximately 45% over the random baseline, with statistical significance ( p<0.05 ) and very large effect sizes (Cohen's d approximate to 2.5 -3.2). Algorithm specialization emerges naturally across network scales, with EFI performing best on small networks and ASC dominating medium to larger configurations. These results indicate that LLM-assisted algorithm synthesis provides a practical and scalable approach for constrained routing optimization in DT-enabled IoT systems.

Accurate and efficient waveform classification is a critical challenge in dense radio frequency environments. While deep learning offers solutions, many models struggle to balance high performance with computational constraints. This paper introduces the Warping State Space Model (WarpSSM), a lightweight and robust framework designed to address this trade-off. Our approach processes spectrograms using two novel components: the Dynamic Adaptive Warp block to mitigate channel-induced geometric distortions, and the Visual State Space Block with Directional Selective Fusion to capture long-range dependencies. Evaluated on a synthetic dataset of 12 waveforms under realistic channel degradation, WarpSSM achieves a state-of-the-art average accuracy of 90.61% . It demonstrates exceptional efficiency, with an inference latency of 0.55 ms from a model of only 42.2K parameters.

Existing deep learning models for spectrum sensing in 5G/6G radar-communication environments face a critical accuracy-efficiency trade-off. To this end, the letter introduces Denoise2Seg, an efficient model that synergistically denoises and segments coexisting 5G, LTE, and radar signals. The core innovation is a multi-task paradigm where an auxiliary denoising task acts as a powerful regularization mechanism, forcing a shared encoder to learn robust, noise-invariant features. This collaborative mechanism is realized through novel architectural modules engineered to perform robust multi-scale analysis and to intelligently arbitrate between clean and noisy feature streams.Through comprehensive experiments, Denoise2Seg demonstrates superior performance, achieving a state-of-the-art mean accuracy of 97.94%, a mean Intersection-over-Union of 96.08%, and a mean F1-score of 98.00% with a compact design of only 4.34M parameters. This represents a significant reduction in complexity compared to other high-performance models, establishing Denoise2Seg as a robust and practical solution for intelligent spectrum sensing in next-generation wireless networks.

In this work, we propose an effective approach to achieve robust covertness in joint radar and multicast-unicast communication using non-orthogonal multiple access (NOMA). Utilizing the same spectrum resource, a dual-functional radar-communication station covertly scans for unauthorized targets while receiving mixed multicast-unicast messages from a base station, including multicast information shared between radar and cell users and unicast covert information intended for cell users. Simultaneously, the unauthorized target monitors the radar's sensing activity to hide its existence before tracking cell users' unicast transmission covertly. NOMA power allocation is considered for both mixed multicast-unicast messages and the transmission between the radar and base station. In this context, we propose enhancing system covertness by first deriving exact closed-form expressions for the minimal detection error probability (DEP) of two-phase monitoring at the unauthorized target and jointly treating them as stringent constraints of system designs, along with the minimal communication rate requirements, when optimizing the power resources of the radar and base station to achieve the mutual information maximization of radar sensing activity. Through in-depth analysis of the DEP in conjunction with intelligent transformations using successive-convex approximation method, we can effectively obtain the near optimal power allocation solution, which is then thoroughly validated through numerical experiments.

Integrated sensing and communication (ISAC) allows the same hardware platform and resources to function sensing and communication simultaneously, which reduces the hardware size and addresses the spectrum congestion concerns. However, the sharing of the hardware and resources of the sensing and communication functions raises resource managements. Traditional optimization approaches are developed based on rigid mathematical models in ISAC systems. However, they face computation complexity and may not achieve the desired performance under the dynamics of the ISAC system environments. Machine learning with ability in learning features/patterns of data as well as approximating mathematical models has recently proposed to effectively solve the complicated ISAC problems. In this survey, we thus provide a comprehensive literature review on applications of learning algorithms for ISAC systems. Particularly, we review learning approaches proposed for emerging issues in ISAC systems, including beamforming designing/tracking, waveform design, spectrum allocation, time allocation, and power allocation, angle of arrival (AoA)/angle of departure (AoD) estimation, signal classification, and security issues. Moreover, we present applications of advanced learning methods for wireless sensing, which is considered to be an emerging sensing service of the next-generation networks. We conclude the survey with highlighting technical issues of learning algorithms and discussing future research directions.

With the growing demand for efficient spectrum utilization in integrated radar-communication (IRC) systems, driven by Internet of Things (IoT) and fifth-generation (5G) advancements, robust waveform classification techniques have become increasingly critical. This article introduces the time-frequency integration deep network (TFINet), a cutting-edge deep learning (DL) architecture designed for waveform classification in spectrally congested environments. TFINet leverages time-frequency representations (TFRs) derived from the smoothing pseudo-Wigner-Ville distribution (SPWVD) to improve feature quality and mitigate cross-term interference, enhancing classification accuracy. The network incorporates two key modules: the dual-temporal frequency extraction (DTFE) and time-frequency selective downsampling (TFSD). The DTFE module improves feature extraction by decoupling time and frequency features through dual-branch processing, while the TFSD module intelligently reduces dimensionality, preserving essential features without compromising performance. These innovations enable TFINet to balance computational efficiency and classification accuracy, enhancing its suitability for resource-constrained edge devices. On a diverse synthetic dataset of 12 waveform types, TFINet achieves 91.38% overall classification accuracy with 59k parameters and 0.328-ms inference time. Compared to existing deep models, TFINet demonstrates superior performance in both accuracy and efficiency, validating its suitability for practical IRC systems.

Healthcare systems are increasingly data-rich but remain fragmented, reactive, and difficult to coordinate across institutions, devices, and clinical workflows. Healthcare 5.0 addresses these limitations through human-centricity, resilience, and symbiotic human-AI collaboration. This survey positions agentic AI, goal-oriented systems capable of contextual perception, persistent memory, multi-step planning, tool use, and bounded action, as the orchestration layer of Healthcare 5.0. We make four contributions: (i) a reference architecture that maps perception, reasoning, planning, action, and memory across device, edge, and cloud tiers; (ii) an autonomy-criticality taxonomy for four healthcare application domains; (iii) a structured analysis of safety, security, privacy, ethics, and regulatory challenges; and (iv) a phased roadmap for responsible clinical deployment. Our analysis shows that Healthcare 5.0 requires not only capable models, but also safety-governed, auditable, and ethically aligned agentic ecosystems.

The rapid proliferation of unmanned aerial vehicles (UAVs) has intensified the need for robust surveillance systems capable of distinguishing drones from biological entities like birds in unpredictable environments. While multispectral vision provides a resilient alternative to uni-modal sensors under adverse weather and lighting, existing architectures often struggle with cross-modal feature alignment and noise-induced spatial distortions. This paper proposes Multispectral Attention Context and Receptive-field Network (MACR-Net), an ultra-lightweight multimodal framework designed for high-precision drone detection. MACR-Net introduces a Global-Local Cross-Scale Interaction (GLCI) module to capture multi-scale semantic context and a Multimodal Spatial Cross-Perception (MSCP) mechanism to adaptively fuse RGB-IR streams while preserving target-specific thermal and structural signatures. Furthermore, we design an improved hybrid neck integrating Coordinate-Aware Attention (CAA) and Receptive Field Deformable (RFD) modules to anchor precise spatial coordinates and mitigate geometric distortions. Experimental results on the benchmark Multimodal Drone Detection Dataset demonstrate that MACR-Net outperforms state-of-the-art models, achieving a peak mAP_50 of 91.13% and a significant mAP_50-95 of 65.77%. Remarkably, the architecture maintains an extremely compact footprint with only 2.77M parameters and 0.77 GFLOPs, establishing an optimal balance between superior detection robustness and real-time feasibility for resource-constrained edge deployment.

This letter introduces an adversarial training approach (ATA), an innovative method designed to enhance the performance of segmentation models in identifying 5G and LTE signals from spectrograms. ATA employs a discriminator model to provide additional guidance, thus refining the segmentation model's learning process by encouraging it to generate prediction masks that not only align with the ground truth but also effectively deceive the discriminator. This adversarial training mechanism significantly optimizes the model's ability to capture complex spectral features. Experimental results demonstrate the broad effectiveness of ATA across state-of-the-art models, including U-Net, U-Net++, SegFormer, DeepLabV3+, SRNet, and PRMNet, achieving substantial performance improvements up to 13.75% in mAcc, 19.83% in mIoU, and 15.03% in mF1. Furthermore, without the need for model architecture modification, ATA preserves the model's computational complexity, thereby offering an efficient solution for intelligent spectrum sensing in wireless communication systems.

The letter proposes an adaptive spectro-temporal robust architecture network (ASTRA-Net), a cutting-edge deep framework for spectrum sensing (SS) in radar and next-generation wireless networks. To overcome the limitations of existing methods in spectrally congested and low signal-to-noise ratio (SNR) environments, ASTRA-Net introduces a dual-path design combining a Dynamic Selective Spectrum Layer (DSSL) for adaptive multi-scale receptive fields, a Pathwise Spatial Layer (PSL) for robust context-aware feature extraction, a Perceptive Spectrum Manifold (PSM) for spectro-temporal decomposition, and an Adaptive Spectrum Distiller (ASD) to enrich global context. Through the coordinated integration of these modules, ASTRA-Net enables precise segmentation of fifth-generation new radio (5G NR), long-term evolution (LTE), and radar signals across diverse noisy environments. Extensive experiments demonstrate that ASTRA-Net achieves 98.65% global accuracy and 97.52% mean IoU with only 6.6M parameters and 2.36 ms inference latency, significantly outperforming recent deep models in both efficiency and robustness. These results indicate that ASTRA-Net represents a highly efficient lightweight architecture capable of jointly balancing accuracy, efficiency, and robustness in wideband spectrum analysis. Open-source code and dataset are released at https://github.com/Phuoc-LongHuynh/ASTRA-Net for reproducibility.

We introduce LDSeCom, a novel lightweight-to-diffusion framework for semantic image communication. LDSeCom addresses bandwidth constraints by developing LSNet, a lightweight, loop-based segmentation model at the sender, and an improved diffusion model guided by our AFM-Net at the receiver. LSNet efficiently compresses images into semantic maps, while AFM-Net’s adaptive feature modulation ensures high-quality image reconstruction. On benchmark datasets, our LSNet achieves competitive accuracy with only 0.5M parameters, while our diffusion model improves image reconstruction quality by up to 28.51% mFID. The framework enables high-fidelity results from semantic maps compressed to 1/80 of the original size, proving its efficiency for bandwidth-constrained scenarios.

In this paper, we investigate the performance of simultaneous transmitting and reflecting reconfigurable intelligent surface (STAR-RIS)-enabled short-packet communication (SPC) systems that employ covert rate-splitting (RS) to facilitate applications of Internet-of-Things (IoT). Under the generalized model of the alpha - eta - kappa - mu fading, we analyze covertness by first deriving a closed-form expression for the warden's detection error probability (DEP) and then deducing the optimal detection threshold at which the DEP is minimized. From this optimal DEP, we guide the choice of the feasible power-allocation (PA) region at which the warden's produced DEP is always beyond the minimal acceptable covertness. On the other hand, we develop mathematical frameworks for evaluating the system's block-error rate (BLER) and the ergodic rate (ER), as well as providing guidelines on how to access their performance limits at high signal-to-noise ratio, especially the diversity orders and ergodic slopes of the users. Furthermore, we also propose to enhance the system performance by jointly optimizing the PA and RS coefficients in order to: 1) minimize the maximum BLER across private users subject to a minimum covertness requirement and 2) maximize the minimum ER across users subject to covertness and decoding constraints. Numerical results confirm the analytical expressions and show that the proposed optimization efficiently tunes the PA and RS coefficients to achieve the BLER and ER fairness objectives.