Along with the increasing of wireless devices and attacks, covert communications face more challenges during spectrum access process. Conventional reinforcement learning (RL) based dynamic spectrum access (DSA) methods often lock onto a single “best” channel with the highest channel capacity, making it easy for attackers to eavesdrop and interfere. To address this issue, this paper proposes a novel aperiodic hopping DSA method for covert communications with Gibbs-sampling inspired RL. Firstly, access-pattern metrics are introduced to characterize the predictability of channel access, including access-pattern exposure based on the Kullback-Leibler (KL) divergence, channel switching rate, and channel residence length. These metrics describe whether a user accesses channels in a balanced manner and whether it persistently stays on the same channel. Then, inspired by Gibbs sampling, we propose a frequency-balancing rejection method with refusing to access overused channels. In such a manner, a user is forced to access multiple channels dynamically, which can reduce the probability of sticking to an attacked channel. Finally, we integrate the proposed Gibbs-sampling-inspired rejection method with RL to form a novel aperiodic hopping DSA framework for covert communications, where the performances of channel capacity and access-pattern exposure are combined together. Simulation results demonstrate that, compared with ε-greedy Q-learning and Softmax-Q Q-learning, the proposed hopping DSA method reduces the average access-pattern exposure by approximately 86.4% and 86.5%, respectively, while maintaining competitive channel capacity. Under the Q-learning-based attacker, it further reduces the average attack hit rate by approximately 47.2% and 52.4%, and increases the average throughput under attack by approximately 44.7% and 51.1%, respectively. By continuously adjusting channel selection, the proposed hopping DSA improves the performance of covert communications.
Collaborative path planning for uncrewed aerial vehicles (UAVs) and ground vehicles is important for low-altitude Internet of vehicles (IoV), where dynamic tasks, heterogeneous resources, and safety constraints challenge conventional routing methods. This paper studies a heterogeneous collaborative delivery system consisting of depot-dispatched UAVs, vehicle-coordinated UAVs, and ground vehicles. The formulation jointly considers dynamic order arrivals and priority escalation, load-dependent energy consumption and replenishment, soft time windows, and airspace feasibility. To solve the resulting dynamic scheduling problem, we propose a priority-biased heterogeneous multi-agent proximal policy optimization (PBH-MAPPO) method. The learned policy operates at the task-assignment level and incorporates priority-aware and cost-aware logit corrections together with priority-weighted advantage estimation, enabling priority information to influence both action selection and policy optimization. Simulations with different fleet scales, very important person (VIP)-task proportions, and obstacle densities show that PBH-MAPPO achieves a favorable trade-off among energy consumption, VIP service quality, and task completion efficiency.
On edge intelligent Internet-of-Things (IoT) platforms, such as unmanned aerial vehicles (UAVs), the need to achieve an effective balance between detection accuracy and energy consumption in object detection models has become increasingly prominent. As a brain-inspired computing paradigm, Spiking Neural Networks (SNNs) have emerged as a promising alternative due to their event-driven and energy-efficient characteristics, yet their potential for UAV small object detection remains largely underexplored. This paper proposes a biologically interpretable spike-driven small object detection model, named BISOD. It follows the YOLO-style backbone-neck-head architecture, while introducing key modules inspired by the functional mechanisms of retinal cells and visual attention in biological vision. These modules are designed to enhance the representation of small object features and alleviate the degradation of weak features caused by the inherent sparsity of SNNs during layer-wise optimization. Extensive experiments are conducted on the VisDrone-DET2019 and UAVDT datasets. The results show that the BISOD model achieves detection accuracy comparable to the baseline YOLOv8m model, while requiring only 2.9M parameters and reducing energy consumption by approximately 90%. Meanwhile, it consistently outperforms existing SNN-based detection models on the evaluated datasets. The proposed method achieves a favorable trade-off among detection accuracy, model lightweightness, and energy efficiency, demonstrating strong feasibility and application potential for deployment on UAV platforms.
In complex agr icultur al scenar ios, image-based obstacle detection suffer s fr om sever e backgr ound inter fer ence (vegetation textur e, shadow occlusion, str ong r eflection), while point clouds exhibit spar se geometr ic r epr esentation and gr ound-dominant distr ibution. Existing multimodal fusion methods lar gely r ely on simple featur e stitching, failing to explicitly lever age geometr ic occupancy and height str uctur es. To addr ess these issues, this paper pr oposes a multimodal detection fr amewor k based on point cloud geometr ic enhancement and object-level geometr ic distillation. We design Agr iVoxel-Mamba to impr ove agr icultur al point cloud r epr esentation via gr ound-obstacle decoupling, distance-adaptive dual-scale modeling, and vegetation suppr ession gating. We intr oduce POGFusion to explicitly inject point cloud geometr ic occupancy and height maps into image featur e space thr ough calibr ated pr ojection, over coming limitations of tr aditional stitching fusion. We fur ther constr uct HGI-CD to per for m geometr y-guided instance-level contr astive distillation with har d sample weighting and categor y pr ototype constr aints. Exper iments on the FieldSAFE dataset demonstr ate that our method achieves a 91.0% r ecall r ate, 97.0% mAP@0.5, and 67.9% mAP@0.5:0.95, outper for ming the single-modal DETRKD baseline by 2.0 per centage points in r ecall and 3.1 per centage points in mAP@0.5:0.95, and compar ed to DINO-R50+ point cloud fusion, the computational cost is r educed by 18.1% while maintaining compar able accur acy. The pr oposed fr amewor k achieves the best accur acy-efficiency/deployment tr ade-off among deployable models, outper for ming single-modal and lightweight multimodal baselines under a compar able compute budget, especially in scenar ios with dense vegetation and occlusion, which validates the effectiveness of our geometr ic enhancement and r efinement mechanism.
Parkinson’s disease progression in elderly populations involves dynamic motor and non-motor symptoms, requiring continuous and interpretable monitoring. This study proposes a CyberTwin-enabled framework integrating deep learning and Explainable Artificial Intelligence (XAI) for early prediction and continuous monitoring of disease progression. The framework enables bidirectional synchronization between patients and their CyberTwin replicas for real-time health-state monitoring and predictive analytics using multimodal IoT wearable data. A hybrid CNN–GRU model extracts spatial features and temporal dependencies, while a SHapley Additive exPlanations (SHAP)-based layer provides interpretable insights into behavioral indicators affecting risk assessment. Experimental results demonstrate a Delay of 2.18 s, Accuracy of 94.72%, Sensitivity of 95.38%, Specificity of 93.84%, F-Measure of 94.58%, and RMSE of 1.34.
