Published papers

Deep learning-based radio frequency fingerprinting (RFF) identification has the potential to enhance the security performance of the physical layer. In recent years, a number of RFF datasets have been proposed to meet the large-scale data requirements for deep learning. However, these datasets are collected from similar channel environments and only contain receiver data. This paper employs different software radio peripherals to generate radio signals. Hence, it is able to adjust the signal's parameters, such as frequency band, modulation style, antenna gain, etc. In this paper, we propose a radio frequency signal dataset based on LTE system and variable channels to more properly characterize the generated signals in the real world. We collect signals at transmitters and receivers to construct the RFF dataset. Moreover, we confirm the dataset's dependability using various machine learning and deep learning methods. The dataset and reproducible code of this paper can be downloaded from GitHub11GitHub link: https://github.com/njuptzsp/XSRPdataset.

High-Altitude Platform Stations (HAPS) promise to extend broadband coverage over wide areas by acting as quasi-stationary aerial base stations. However, wind-induced movements cause antenna misalignment and unbalanced user distributions, leading to throughput degradation. This work presents a low-complexity, equal clustering (EC) -aided deep Q-network (DQN) algorithm for dynamic antenna control in HAPS networks. The proposed EC algorithm partitions users into balanced clusters by identifying low-density sectors in a single pass, avoiding the iterative refinements of other clustering algorithms such as K-Means. A DQN then fine-tunes the antenna tilt angles and beam widths based on the EC results to maximize the throughput of low-throughput users. The proposed method’s performance is evaluated through simulations under rotation and shift scenarios and benchmarked against three Reinforcement Learning (RL)-based approaches and a K-Means clustering baseline. Compared with these methods, the EC-aided DQN delivers higher throughput, especially for users in the low-throughput tail, while reducing computational complexity. For instance, the percentage of low-throughput users was reduced by over 10% in the case of 5 km shift in the regions of Osaka and Sendai, compared to K-means. Similarly, in the case of 30° rotation, the percentage is reduced by 4.1% and 6.5% for the same regions. Statistical significance tests confirm that the improvements are not due to random simulation noise.

This paper presents a radar-based heart rate monitoring and Inter-beat interval (IBI) estimation. Traditional IBI estimation relies on peak detection combined with signal processing techniques. Our approach uses signal reconstruction with a neural network, which significantly improves the accuracy and robustness of IBI estimation compared to conventional methods. By utilizing multi-channel input data, our approach effectively mitigates various sources of noise and interference, resulting in highly accurate and reliable IBI predictions. Instead of relying solely on peak detection, we use the U-net architecture and cross-channel attention mechanism to reconstruct the triangular waveform generated from the ground truth ECG signals. The incorporation of multi-channel input and channel-wise attention mechanism improves our model’s ability to discriminate and emphasize critical features from different input channels, further improving IBI estimation accuracy. To evaluate the accuracy and robustness of our proposed method, we trained our model on the open-source dataset [1], and then performed leave-one-subject-out validation, which ensures our approach to be evaluated only on unseen independent subjects. We have achieved a Root Mean Square Error (RMSE) of 26.7 ms for the IBI of each heart-beat. Our results demonstrate the transformative potential of adopting signal reconstruction methods supported by state-of-the-art deep learning techniques. This shift in perspective promises more accurate and robust heart rate and IBI estimation, opening new avenues for improving the accuracy and reliability of cardiac monitoring systems.

We propose a neural network-based framework for radar-based heart rate monitoring and inter-beat interval (IBI) estimation that reconstructs the underlying cardiac waveform, rather than relying on Conventional peak detection methods. Conventional approaches typically estimate IBI by detecting peaks in radar signals after extensive signal processing, which can be sensitive to noise and transient signal degradation. Building on our previous work, which utilized a U-Net architecture with cross-channel attention, we introduce a enhanced model by incorporating Bidirectional Long Short-Term Memory (BiLSTM) layers and temporal self-attention into the bottleneck of the U-Net. This modification allows the network to explicitly model temporal dependencies and heartbeat periodicity, improving resilience to transient input degradation and physiological variability. The results are evaluated on a public radar-Electrocardiogram(ECG) dataset under Valsalva, Tilt Up, and Tilt Down conditions, our model achieves an average IBI estimation error of 22.54 ms Root Mean Square Error(RMSE), which outperforming both conventional peak-based methods and our prior U-Net-based design. These results demonstrate improved temporal generalization and signal stability, advancing the feasibility of radar-based cardiac monitoring in clinical and home settings.

MIMO FMCW Radar, a non-contact heart rate (HR) monitoring technology, has emerged as one of the superior alternatives to contact-based sensors, providing precise HR estimation without the drawbacks of discomfort or privacy concerns, and excelling in diverse environmental conditions. Recent HR estimation studies using conventional methods have achieved high accuracy but face challenges with lengthy processing times and sensitivity to experimental conditions, affecting real-time application and robustness. A previous Deep learning (DL)-based approach improves on these aspects but is limited by the range resolution of SISO FMCW Radar and the requirement for close positioning of subjects. Additionally, this method has inability to utilize contextual information from adjacent time windows restricts its effectiveness for time series analysis. Thus, our proposed method combines a Curve-Length (CL) approach with a DL rough-to-fine model, addressing the limitations of previous studies and improving HR estimation accuracy and robustness. This integrated process begins with human location detection through a CL-based approach and is followed by a two-phase rough-to-fine HR estimation. The experimental results show significant improvement over the conventional methods.

Synthetic Aperture Radar (SAR) can operate under all-weather, all-day conditions, playing a crucial role in regional change detection (CD). However, due to its unique imaging principles, SAR images contain significant speckle noise and blurred boundary and detail features, which reduces the detection accuracy and leads to missed detection and false detection. To address these issues, this paper proposes a SAR image CD method based on saliency region guidance and Scale-Invariant Feature Transform (SIFT) keypoint extraction to reduce the interference of speckle noise. First, a saliency region guidance method is introduced to analyze the saliency of local features in SAR images, extracting potentially changed regions and reducing the interference of speckle noise. Second, the SIFT is employed to extract keypoints in regions significantly different from the background in the difference map, leveraging its robustness to speckle noise. By extracting keypoints, the approximate location and extent of the changed regions are determined. These are, then, fused with the saliency region information, enhancing the saliency weights of pixels around keypoints for more extraction of change regions. Finally, a Vision Transformer (ViT) detection network is used for SAR image CD, utilizing the combined saliency information from the original saliency map and SIFT keypoints. This approach effectively integrates SIFT's stable description of local features with ViT's modeling capability for global features, improving the model's accuracy and robustness.

Federated learning (FL) offers significant advantages in preserving data privacy and enhancing communication efficiency, making it especially suitable for Internet of Things (IoT) networks supported by uncrewed aerial vehicles (UAVs). However, most existing FL approaches rely on assumptions of uniformly distributed, well-labeled, and large-scale datasets—conditions that rarely hold in practical UAV-based IoT scenarios. These environments typically feature small-scale, nonindependent and identically distributed (non-IID), and dynamically changing data. To address these challenges, we propose a novel self-supervised federated unsupervised learning (FUL) framework tailored for UAV-assisted IoT systems. The proposed framework comprises three key components: 1) a realistic UAV data collection model that considers limited onboard storage and mobility constraints; 2) a robust local training strategy that incorporates self-supervised regularization and a centered kernel alignment (CKA)-based similarity loss to mitigate the effects of data heterogeneity and rapid distribution shifts; and 3) an importance-aware hybrid normalized aggregation method at the global server, which leverages model divergence-based metrics to evaluate local model reliability and integrates both current and historical gradient information for stable model updates. Experimental results demonstrate that our framework achieves classification accuracies of 30.5%, 62.8%, and 70.5% under memory constraints of 500, 1000, and 2000 samples, respectively—outperforming the best baseline by 18.1%, 26.5%, and 9.1% under the same conditions. These results highlight the effectiveness of the proposed FUL framework in handling data heterogeneity and dynamic sample variations inherent in realistic UAV-enabled IoT applications.

Semantic communication (SC) has emerged as a promising paradigm to address the bandwidth limitations of traditional wireless communication systems by transmitting only the essential meaning of data. This paper investigates an advanced SC framework for vehicular communication, employing diverse feature extraction techniques to encode multi-modal information, such as textual descriptions, object poses, semantic segmentation, and sketches, into compact semantic representations. These semantic encoders are evaluated based on their output size, faithfulness of reconstruction, and resilience to data loss. The proposed system model considers vehicular communication scenarios where vehicles transmit important information extracted from camera-collected data to other vehicles and road users, in bandwidth-constrained environments. Simulation results show the effectiveness of the proposed SC framework, with a reconstruction performance reaching 17.24 in Fréchet Inception distance (FID) and a an RMSE equal to 0.029 between the transmitted image and the reconstructed one. This performance is achieved while the data saving between the size of the original image and the transmitted semantics is equal to 92.25%.

Malware traffic classification (MTC) technology plays a crucial role in cyber security by serving as the first step of network intrusion detection system. Recently, deep learning has been successfully introduced into the field of network intrusion detection. However, its performance relies heavily on the quantity and quality of labeled samples, which are often difficult to obtain in real-world applications. To address these challenges, we propose semi-supervised MTC method based on consistency regularization and pseudo-labeling. This method utilizes unlabeled malware traffic data to improve the classification performance of the model. Specifically, we train the model using labeled malware traffic data. Then, weak data augmentation is performed on unlabeled traffic data and pseudo labels are generated through the model. Finally, the model predicts unlabeled data after strong data augmentation and compares it with pseudo labels. Model parameters are updated using losses from labeled and unlabeled samples. Experimental results demonstrate that our proposed method outperforms other classical classification approaches, including ladder networks, virtual adversarial training (VAT), convolutional neural network (CNN), and support vector machine (SVM), when the ratios of labeled samples range from 1% to 5%.

Classification of ships in synthetic aperture radar (SAR) images plays a critical role in maritime search and rescue as well as national defense security. Over the past few years, the swift advancement rapid of deep learning has opened up innovative technological directions for precise classification of SAR ship images. However, the unique imaging mechanism of SAR images, resulting in a lack of ship samples and insufficient feature information, continues to make accurate classification of SAR ships challenging. In this study, we propose an improved vision transformer (ViT) method using a full attention (FA) mechanism, incorporating channel attention after Self-Attention, dynamically selecting channels through reweighting to enhance comprehensive feature representation. To address the issue of insufficient training samples, we designed an image reconstruction module based on SRGAN, which generates high-resolution and detailed image data through adversarial learning, effectively compensating for the sample quantity in the dataset. Ultimately, the results section demonstrates that the method introduced delivers enhanced classification accuracy relative to current methods.

In recent years, non-contact Blood Pressure (BP) measurement has been attracting attention for measuring our health status in daily life. A Doppler radar can observe pulse waves caused by chest wall displacement due to heartbeat. BP can be estimated by constructing a BP estimation model using BP-related features obtained from the pulse wave. However, compared to when modeling for each subject, the BP esti-mation accuracy deteriorates significantly when modeling with multiple subjects including the testing subject. To deal with this limitation, BP category classification has been introduced into PhotoPlethysmoGraphy (PPG)-based BP estimation. In this paper, we develop a Doppler radar-based BP estimation method based on BP category classification. In the proposed method, the pulse waves extracted from a Doppler radar are classified into three categories, 'Low BP', 'Normal BP', and 'High BP' by k-Nearest Neighbor (kNN) based on the features that correlate with BP. The SBP estimation model is trained for each BP category. After the BP category classification, SBP is then estimated by using the model corresponding to the classified BP category. The experimental results showed that the proposed method with BP category classification estimated SBP accurately, compared to without BP category classification.

On-device fine-tuning of large language models (LLMs) has attracted a lot of attention because of its tailoring personalized models while retaining user data locally on the mobile device. However, it faces significant challenges due to prohibitive memory requirements and slow training speeds. In this paper, we propose MobiLLM, a novel scheme enabling memory-efficient LLM fine-tuning on a single mobile device via server-assisted side-tuning. Particularly, MobiLLM strategically offloads backpropagation computations to an edge server while allowing the resource-constrained mobile device to retain merely a pretrained backbone model with frozen parameters during finetuning. It constructs a backpropagation bypass via parallel adapters decoupled from the backbone. During forward propagation, the device employs low bitwidth quantization for transmitting intermediate activations to the server to reduce communication overhead. The advantage of MobiLLM lies in: 1) confining training data strictly to the mobile device, and 2) eliminating on-device backpropagation while overlapping local computations with server execution. Collectively, MobiLLM ensures the data never leaves the local mobile device while significantly reducing mobile memory and computational burdens. We implement MobiLLM on several popular mobile devices, including NVIDIA Jetson Xavier NX and CPU-only laptops. Extensive experimental results demonstrate that MobiLLM can enable a resource-constrained mobile device to fine-tune billion-sized LLMs, achieving up to 4x memory reduction and 2.3x faster convergence as compared to state-of-the-art baselines.

Beamforming plays a key role in improving the spectrum utilization efficiency of multi-antenna systems. However, we observe that (i) conventional beam prediction solutions suffer from high model training overhead and computational latency and thus cannot adapt quickly to changing wireless environments, and (ii) deep-learning-based beamforming may face the risk of catastrophic oblivion in dynamically changing environments, which can significantly degrade system performance. Inspired by the above challenges, we propose a continuous-learning-inspired beam prediction model for fast beamforming adaptation in dynamic downlink millimeter-wave (mmWave) communications. More specifically, we develop a meta-empirical replay (MER)-based beam prediction model. It combines empirical replay and optimization-based meta-learning. This approach optimizes the trade-offs between transmission and interference in dynamic environments, enabling effective fast beamforming adaptation. Finally, the high-performance gains brought by the proposed model in dynamic communication environments are verified through simulations. The simulation results show that our proposed model not only maintains a high-performance memory for old tasks but also adapts quickly to new tasks.

Radio Frequency Fingerprinting (RFF) is a critical technology for enhancing physical-layer security by leveraging the unique RF characteristics of hardware, enabling authentication and anti-counterfeiting for wireless communication devices. In recent years, Deep Learning (DL) has been extensively applied in RFF, significantly improving identification accuracy and efficiency. However, DL- based RFF methods still encounter challenges regarding robustness, particularly in cross-channel scenarios. To address these challenges, we propose a channel-robust RFF method based on a Multi-Scale Convolutional Attention Network (MSCAN) with Random Overlay Augmentation (ROA). Specifically, MSCAN extracts and fuses features at different scales, allowing it to capture more comprehensive signal characteristics. Additionally, ROA is a combinatorial data augmentation (DA) strategy designed to simulate diverse characteristics of wireless propagation environments, thereby enhancing the adaptability and robustness of RFF in complex channel conditions. Experiments conducted on the ORACLE dataset demonstrate that our proposed method achieves over 92 % accuracy in cross-channel scenarios, outperforming the previously proposed DA strategy. The codes will be published in GitHub11https://github.com/BeechburgPieStar/SDG-for-Robust-SEI

Wi-Fi channel state information (CSI)-based human activity recognition (HAR) leverages rich channel propagation characteristics to enable nonintrusive, privacy-preserving, and device-free sensing. In multiuser wireless environments, however, HAR faces significant challenges, including multipath interference, signal overlap, label ambiguity, cross-domain channel variability, and constraints imposed by real-time deployment on resource-limited edge devices. This article presents multiuser enhanced convolutional network (MuECNet), a lightweight and modular deep learning framework designed to operate under realistic multiuser, multiactivity CSI sensing conditions. MuECNet integrates three key components: 1) an enhanced convolutional encoding module (ECEM) for fine-grained temporal-spectral feature extraction that preserves channel propagation signatures; 2) a branch-wise feature normalization (BFN) module for user-specific channel representation learning; and 3) an adaptive decision module for multilabel activity inference. To improve robustness under diverse and dynamic channel conditions, we introduce MixUp-based data augmentation to emulate activity overlap and reduce label ambiguity. Evaluations on the WiMANS dataset show that MuECNet achieves 64.46% (2.4 GHz) and 64.15% (5 GHz) accuracy with only 1.19-M parameters, 1.74-G FLOPs, and 0.28-s inference latency, outperforming baseline models such as ABLSTM and THAT while reducing model size by up to 75%. Ablation studies confirm the contribution of each module, with accuracy drops of up to 6.41% when removed. These results demonstrate that MuECNet provides a robust and communication-efficient solution for integrating CSI-based sensing into future wireless networks and Internet of Things (IoT) systems.

Semantic communication plays a crucial role in human interactions, allowing for the exchange of complex ideas and concepts. In this paper, we introduce a novel approach to semantic communication leveraging image generative Artificial Intelligence (AI) models, specifically stable diffusion models. Unlike conventional works, our system enables the transmission of images through a physical channel by transforming them into multi-dimensional semantic representations consisting of text descriptions, low-resolution sketches, and pose information. At the receiver's end, these semantic representations are used to reconstruct the original image using a trained stable diffusion model. The benefits of our approach include reduced transmission bandwidth requirements, flexibility in reconstruction styles, adaptability to multiple receivers' preferences, and the ability to omit unwanted image elements. We present preliminary results demonstrating the feasibility and effectiveness of our method. The similarity score between the transmitted images and reconstructed ones reach values ranging between 0.015 and 0.029 in Root Mean Square Error (RMSE) and between 0.993 and 0.998 using a Siamese network.

Multi-dimensional spectrum prediction is essential for spectrum sharing and dynamic spectrum access (DSA), tack-ling spectrum scarcity and improving wireless communication. Traditional methods often use machine learning (ML), which requires manual feature extraction, or deep learning (DL), which demands high computational resources. This paper proposes a lightweight multi-dimensional spectrum prediction model using an adaptive broad learning network (ABLN). The model employs a sliding window to preprocess data and establishes input layers using randomly generated feature and enhancement nodes. The weights of broad learning are determined by solving the pseudo-inverse, and the structure is incrementally extended without retraining, reducing computational complexity. An adaptive node increment module optimizes hyperparameters efficiently. Experimental results demonstrate that ABLN reduces computational overhead while maintaining robust prediction performance across various scenarios.

In this paper, we propose an unmanned aerial vehicles (UAVs) assisted emergency caching framework based on non-orthogonal multiple access (NOMA), involving a primary UAV and several secondary UAVs. The primary UAV obtains rescue information from a remote base station (BS) and sends it to secondary UAVs hovering at designated locations to serve ground users. Specifically, a multi-objective optimization problem is formulated to maximize the number of files that can be successfully decoded in the primary UAV and minimize the number of user clusters for secondary UAVs. The multi-objective optimization problem is decomposed into three sub-problems: primary UAV deployment, power allocation, and secondary UAV deployment. First, a primary UAV deployment scheme is proposed to maximize the number of received files subject to the constraint of the demand for the line-of-sight (LoS) link probability between the primary UAV and the ground users. Subsequently, to achieve the same goal, a closed-form solution for the NOMA power allocation factor of the files is derived. Finally, due to the limited range of services of the secondary UAVs, we propose a secondary UAV deployment scheme to minimize the number of their flights, which can save energy and reduce user latency. Extensive numerical results demonstrate that the proposed algorithms outperform the random scheme and the worst scheme.

Change detection based on synthetic aperture radar (SAR) images is a challenging task in the field of remote sensing image analysis due to the influence of noise and the lack of labeled data. In this paper, we propose a new unsupervised change detection algorithm based on deep learning, which explores the spatial and frequency domain features of SAR images in parallel to improve detection performance. Our proposed method first obtains pseudo-labels by clustering and then combines them with neural networks for unsupervised detection. To reduce the impact of noise and improve sensitivity to changes, we integrate an attention mechanism (AM) into the network. We also use complementary features to integrate the spatial and frequency domain features. These complementary features include a multi-regional feature weighted by channel-spatial AM and a deep feature filtered out by a gated linear unit (GLU). Experimental results demonstrate that the proposed method improves the detection accuracy.

Specific emitter identification (SEI) extracts the fingerprint characteristics of emitters according to the subtle differences of transmitted signals, to distinguish different emitter individuals and prevent unauthorized network access. Deep learning (DL) based SEI methods have been proposed to achieve a good identification performance in recent years. However, these methods highly rely on expert experience to design network structures. These hand-designed fixed network structures lack flexibility, which often leads to insufficient model generalization. Neural architecture search (NAS) can be seen as a subdomain of automatic machine learning (AutoML), which can automatically adjust network structure and parameters according to a specific task. In this paper, we propose a neural architecture search-based SEI method, which can achieve an efficient search of the architecture with the use of a gradient descent algorithm. Experimental results show that the proposed NASEI method both improves the accuracy and reduces the parameter quantity when compared with state-of-the-art methods. Code available at https://github.com/huangyuxuan11/NASEI.git.