Published papers

Conventional multi-unmanned aerial vehicle (UAV) assisted non-orthogonal multiple access (NOMA) wireless caching networks (WCNs) usually operate in a distributed and non-collaborative manner, where each UAV serves users independently without coordination or relay support. When UAVs move beyond the communication range of ground base stations (BSs), backhaul disruption occurs, leading to high user latency and limited system scalability. To address these issues, we propose a multi-layer UAV-assisted NOMA WCN architecture, where a primary UAV (PUAV) communicates with the BS and cooperates with multiple secondary UAVs (SUAVs). The PUAV not only acts as a control and coordination hub but also serves as a relay for content transmission to SUAVs when necessary. To minimize user transmission latency, we propose a joint iterative algorithm that integrates user clustering, user pairing, power allocation, and UAV deployment. First, we develop an Advanced Balanced K-Means++ (ABKM) algorithm to ensure that each cluster contains a balanced number of users and to reduce the distance between users and their serving SUAVs. Next, we derive the NOMA power allocation factor that minimizes user transmission latency, ensuring efficient resource distribution among all paired users. Furthermore, we analyze the impact of PUAV and SUAV placement on user latency and propose a two-stage particle swarm optimization (PSO)-based algorithm to iteratively optimize the deployment of all UAVs. Finally, the user pairs and power allocation are jointly optimized based on the updated deployment of all UAVs to further reduce user latency. Simulation results show that, compared with a single-layer UAV architecture and benchmark schemes, the proposed multi-layer design with joint optimization achieves lower user latency. Additionally, comparisons with the optimal power allocation search method confirm the validity of the derived NOMA power allocation factor.

With the rapid proliferation of uncrewed aerial vehicles (UAVs) in civilian and industrial applications, the risk of malicious or unauthorized UAV use has become a critical security concern. Existing machine learning (ML)-based UAV recognition methods offer a certain degree of interpretability, but their performance is often limited in complex environments and across diverse UAV types. In contrast, deep learning (DL)-based methods exhibit strong representation capability, yet they generally lack physical interpretability. To address this issue, we propose an interpretable UAV recognition framework, termed frequency-aware network for UAV recognition (FANet-UAV), which performs coarse-to-fine feature learning in the frequency domain. Specifically, a multiplication filter module (MFM) is first designed to capture coarse-grained spectral patterns by exploiting multi-mode and multi-scale frequency characteristics of UAV signals. Based on these coarse representations, a convolutional neural network (CNN) is further employed to extract fine-grained discriminative features for accurate classification. Experimental results on two public UAV datasets demonstrate the effectiveness of the proposed method. In particular, FANet-UAV improves the recognition accuracy from 90.45% to 96.82% on DroneRFa and from 94.15% to 98.83% on DroneRF. Moreover, visualization results and channel-wise SHAP analysis provide both pre-hoc and post-hoc interpretability, revealing that FANet-UAV mainly relies on flight control signal (FCS) features for decision-making, while video transmission signal (VTS) features contribute less to the final recognition results.

High Altitude Platform Station (HAPS) systems require dynamic beamforming to handle platform mobility and uneven user distributions. However, traditional reinforcement learning (RL) and evolutionary algorithms run the risk of degrading throughput during exploration. We propose a model-based deep RL (DRL) framework that pre-trains an environment model to map antenna parameters and HAPS states to SINR distributions. This simulated feedback allows efficient parameter exploration without reducing average throughput below its initial level. Simulation results show that our method consistently maintains higher throughput than conventional RL approaches and swiftly adapts to network changes.

Reconfigurable intelligent surface (RIS)-assisted communication has emerged as a critical enabler for sixth-generation (6G) wireless networks. In particular, the active RIS architecture, by virtue of its signal amplification capability, overcomes the double-fading bottleneck inherent to passive RIS and provides a new avenue for secure physical-layer transmission. However, existing studies have not adequately addressed the secrecy performance of RIS in delay-constrained quality-of-service (QoS) scenarios. This letter investigates the problem of maximization of effective secrecy rate (ESR) in an active RIS-assisted multiple-input single-output (MISO) wiretap channel, subject to both transmit power and active RIS power constraints, under delay-limited QoS requirements. We tackle the nonconvex problem using an alternating optimization (AO)-based iterative algorithm. An analysis of the algorithmic complexity is also presented. Experiment results verify the performance gains achieved by the optimized design.

Dynamic Simultaneous Localization and Mapping (SLAM) remains a challenging problem in robotics, particularly due to the interference of moving objects with feature-based localization and mapping. In this paper, we introduce a novel dynamic object-aware SLAM system built upon ORB-SLAM3, which integrates a computationally efficient Vision Mamba (VMamba) model with an advanced dynamic object filtering strategy. The proposed VMamba model, comprising only 8.8 million parameters, achieves high-precision object detection with minimal computational overhead, thereby enabling real-time operation on embedded platforms. A key contribution of this work is the development of an improved dynamic object filtering strategy that selectively retains static features within dynamic object bounding boxes. By leveraging Harris corners, optical flow, and epipolar geometry, our approach effectively eliminates dynamic outliers while preserving essential static scene information—a significant advancement over conventional methods that indiscriminately remove features. Experimental evaluations on the TUM Dynamic and BONN datasets demonstrate that our method attains superior Absolute Trajectory Error (ATE) performance compared to state-of-the-art dynamic SLAM approaches, including ORB-SLAM3, DynaSLAM, and YOLO-SLAM. Furthermore, real-time deployment on an NVIDIA Jetson Orin NX 16 GB platform incurs only a 6.45% reduction in Frames Per Second (FPS) relative to YOLO-SLAM, while achieving a 28.68% improvement in average ATE. These results underscore the computational efficiency of VMamba and the robust real-time capabilities of our SLAM system. Our study validates that integrating a lightweight, high-precision object detection model with an effective dynamic object filtering strategy can significantly enhance the robustness, accuracy, and real-time performance of dynamic SLAM, even on resource-constrained platforms. Future work will focus on further optimizing VMamba and extending the dynamic filtering techniques to better accommodate increasingly complex dynamic environments, thereby broadening its applicability in real-world robotic systems.

As generative audio models are rapidly evolving, AIgenerated audios increasingly raise concerns about copyright infringement and misinformation spread. Audio watermarking, as a proactive defense, can embed secret messages into audio for copyright protection and source verification. However, current neural audio watermarking methods focus primarily on the imperceptibility and robustness of watermarking, while ignoring its vulnerability to security attacks. In this paper, we develop a simple yet powerful attack: the overwriting attack that overwrites the legitimate audio watermark with a forged one and makes the original legitimate watermark undetectable. Based on the audio watermarking information that the adversary has, we propose three categories of overwriting attacks, i.e., white-box, gray-box, and black-box attacks. We also thoroughly evaluate the proposed attacks on stateof-the-art neural audio watermarking methods. Experimental results demonstrate that the proposed overwriting attacks can effectively compromise existing watermarking schemes across various settings and achieve a nearly 100% attack success rate. The practicality and effectiveness of the proposed overwriting attacks expose security flaws in existing neural audio watermarking systems, underscoring the need to enhance security in future audio watermarking designs.

Human activity recognition (HAR) has received intensely attention in many applications, such as healthcare, human-computer interaction, and smart home. Existing HAR methods based on deep learning (DL) have been proposed in the last several years. However, these DL-based HAR methods are hard to balance between performance and cost, which truly limited the applications in practical scenarios. To solve this problem, this paper proposes a smartphone-aided HAR method using the residual multi-layer perceptron (Res-MLP). It composes of two linear layers and Gaussian error linear unit (GELU) activation function, and obtains Res-MLP network through residual. Experimental results show that the proposed HAR method can achieve a high classification accuracy of 96.72% based on the public UCI HAR dataset.

Twitter has skyrocketed over the past few years and has become a major social media platform. At the same time, the number of social robots on Twitter has also increased significantly. These bot accounts imitate the speeches of normal users to manipulate public opinions, affect the normal communication of users. Therefore, bot account detection came into being. Despite extensive research efforts, bots on Twitter are still evolving to evade detection. Most of the current bot detection methods have a single structure and cannot detect and identify different types of bot accounts well. In this paper, we propose a new system for social robot detection that uses a RoBERTa (Robustly Optimized Bidirectional Encoder Representations from Transformers Pretraining Approach) classifier and a random forest regressor with similarity analysis. In particular, the system considers the similarity of tweets and uses a voting system in addition to a set of features extracted from the user profile information and the tweets themselves. We conduct experiments using the largest dataset of bots available and show that the accuracy of our system is up to 0.8588, which is higher than that of all the other baseline methods.

In future multi-tier cellular networks, cognitive radio (CR) compatible with device-to-de-vice (D2D) communication can be an aid to enhance system spectral efficiency (SE) and energy efficiency (EE). Users in proximity can establish a direct connection with D2D communication and by-pass the base stations (BSs), thereby offloading the network infrastructure and providing EE im-provement. We use stochastic geometry to model and analyze cognitive D2D communication underlying a multi-tier/multi-channel cellular network where the D2D transmitters are capable of harvesting RF energy from ambient interference resulting from simultaneous cellular downlink trans-missions. For further improvement in EE, small cells (SCs) can be put into a power-saving mode by specifying a load-dependent transmission power coefficient (TPC) for SC BSs. In addition, to con-sider practical D2D communication scenarios, we propose a wireless video sharing framework where cache-enabled users can store and exchange popular video files through D2D communica-tion. We investigate the potential effects of the TPC and the introduced D2D layer on the network EE and SE. We will also observe that the energy-harvesting CR-based D2D communication network design will not only ease the spectrum shortage problem but will also result in a greener network thanks to its reliance on ambient energies.

The Doppler ultrasound (DUS) transducer has been widely used for fetal heart rate (FHR) monitoring. However, the fetal DUS signals from the transducers can be corrupted by several interference sources such as maternal and fetal movements, which makes FHR estimation using fetal DUS signals challenging. Fetal DUS signal quality assessment (SQA) can help to remove or interpolate unreliable FHRs estimated from noisy signals to improve the accuracy of FHR estimation. There are some existing approaches for fetal DUS SQA, and most of these approaches with high accuracy are based on supervised learning-based algorithms and human-defined properties. Nonetheless, the fetal DUS datasets with quality-level annotations are limited, and human-defined properties place a limitation on mining more deep information related to signal quality in fetal DUS signals. In this paper, we propose an unsupervised representation learning-based fetal DUS SQA for the improvement of FHR estimation performance. We firstly learn representations of pre-processed fetal DUS data from variational autoencoder (VAE) and then combine these representations as one signal quality index (SQI) using a self-organizing map (SOM). Finally, we apply the combined SQI and a Kalman filter (KF) to estimate fetal RR intervals (FRRI) for reducing the errors of FHR estimation. The experimental results showed that our proposed method could reduce the averaged root mean squared error (RMSE) of FRRI and averaged absolute error (AAE) of FHR.

Transfer learning-based methods for Channel State Information (CSI) feedback can achieve a good CSI reconstruction performance with small amounts of computation time and cost. However, for higher efficacy, such methods need to fine-tune a properly selected source channel model each time. This is because fine-tuning a model trained on drastically different data will suffer from slow convergence and non-optimal performance, which defies the purpose of transfer-learning. In this paper, we use a one-stage algorithm utilizing k-nearest neighbors (k-NN) and Deep Neural Network (DNN) methods to classify the downlink Clustered Delay Line (CDL) channels into one of 5 channel models with the help of a selected set of features. Based on the classification results, the appropriate source channel model for fine-tuning is selected, and the model is fine-tuned on the received data. To demonstrate the usefulness of our method, we compare its results to those of existing methods. Through simulation, we show that the overall classification accuracy for 5 CDL channel models using the one-step algorithm is 80.73%. Upon classification, and by selecting the proper source model, our method can reach a better normalized mean square error (NMSE) of CSI feedback performance in a shorter time and with fewer samples.