Internet Technology Letters最新文献

In this era, large-scale Wireless Sensor Networks (WSNs) provide high Quality of Service (QoS) with energy awareness and scalability. Specifically, existing clustering protocols lack adaptability to evolving networks owing to static cluster formation, data-centric greedy protocols, and their handling of clustering aggregators. To resolve these issues in WSNs, a transformer-based hybrid protocol using a swin transformer for clustering and a decision transformer for energy-efficient routing is proposed, which is called a Swin-Decision Transformer (SDT). Hence, this research employs five main modules. The first module incorporates a Gated Graph Convolutional Network (GGCN) for distributed cluster formation that utilizes the topology of each sensor node. The second module utilizes a Swin Transformer to select a context-aware Cluster Head (CH), scoring on attention for energy, Signal to Noise Ratio (SNR), and load. Furthermore, the third module uses a transformer-based path planner to route the Mobile Data Collector (MDC) to an optimal route for CH visitation. Furthermore, the Decision Transformer is utilized to dynamically route the CH rather than define a path to the Base Station (BS), which assists in the optimization of CH-aided positioning based on learning QoS-optimized trajectories. Ultimately, reinforcement learning is employed in a single change loop to continuously adjust the model parameters. Thus, the simulations demonstrate that SDT improves residual energy, reduces latency, and enhances throughput compared with MDC protocols, which reduce routing overhead by up to 40%.

This paper presents a novel approach for defect detection in infrastructure using unmanned aerial vehicles (UAVs) integrated with Building Information Modeling (BIM) and edge computing. The proposed system leverages BIM's rich geometric and semantic data to guide UAV flight paths and defect localization, while edge computing is employed to perform real-time defect detection onboard the UAV. The methodology includes four key components: BIM-driven UAV flight planning, lightweight onboard defect detection via edge-enabled deep learning models, coordinate transformation for BIM alignment, and the instantiation of detected defects into the BIM model for lifecycle management. Experimental results demonstrate that the proposed BIM-UAV-Edge framework outperforms traditional image processing and deep learning-based approaches in terms of accuracy, latency, and system efficiency. With an F1-score of 0.92 compared to 0.87, a processing latency of 110 ms versus 300 ms, and network bandwidth efficiency of 4 Mbps against 30 Mbps, this integrated solution significantly reduces data transmission and improves real-time feedback, offering a robust framework for large-scale infrastructure monitoring and predictive maintenance.

The advance of Internet-of-Things (IoT) devices in smart grids has enabled fine-grained data collection for electric load prediction but raises concerns of data privacy and security. Traditional methods train a prediction model in a centralized way, which creates a vulnerability to privacy leakage and cyber-attacks. In addition, the collected data by IoT devices often suffer from measurement noise and imbalanced patterns, which may degrade the performance of forecasting models. To address these challenges, we propose in this paper a novel approach, called Federated Electric Load Distribution Learning (FELDL), which integrates Label Distribution Learning (LDL) within a Federated Learning (FL) paradigm. FELDL generates a load distribution for each data point to model data noise and imbalance, and learns such distributions in an FL framework without centralizing data from multiple IoT devices. Finally, we evaluate the performance of FELDL on real-world power consumption datasets, and the experimental results demonstrate that FELDL achieves competitive performance against the comparing methods. Overall, FELDL provides an effective and secure solution for accurate load prediction in IoT-enabled smart grids.

In the Internet of Vehicles (IoV), a critical subset of the Internet of Things (IoT), autonomous vehicles function as sophisticated mobile edge nodes that must process vast streams of sensor data in real-time. Accurate pedestrian detection is a safety-critical task for these nodes, yet delivering high recall for tiny and distant pedestrians on milliwatt-class IoT hardware remains a significant challenge. We introduce a TinyML-oriented detector that couples (i) a sparsity-aware, pruned real-time detection architecture, (ii) an ultra-light channel–spatial attention block, and (iii) a progressive pruning and mixed-precision quantization pipeline co-optimized for resource-constrained IoT-edge platforms. The final network is 86% smaller than its uncompressed baseline and runs at 12 FPS within a 5.3 W power budget on a representative embedded IoT device. Across CityPersons, Caltech-USA, and KITTI datasets, our model consistently outperforms recent lightweight detectors, achieving a 15 percentage-point reduction in the log-average miss rate for the challenging case of pedestrians under 50 pixels. The results demonstrate that our hardware-aware, TinyML approach enables reliable, real-time pedestrian perception on low-power automotive IoT-edge nodes.

This paper proposes a 6G-driven federated edge intelligence framework for multi-center stroke lesion segmentation. Lightweight MobileStroke-U-Net is deployed at the edge of each participating hospital for distributed MRI segmentation, enabling federated learning without data sharing. Gradient adaptive weighted aggregation (GradAdapt) is used at the Center Server layer to alleviate heterogeneous distribution offset across multiple centers. The global model is trained in the cloud and combined with homomorphic encryption and blockchain auditing to achieve end-to-end privacy protection. Experiments are conducted on two major multi-center datasets, ATLASv2.0 and ISLES22: Compared with the best single model, FL-MobileStroke-U-Net improves Dice by 1%–2% on both datasets. FL-MobileStroke-U-Net improves Dice from 0.6458 to 0.6611 on ATLAS v2.0, and the 95% Hausdorff distance changes from 22.9573 to 21.2032; on ISLES22, Dice increases from 0.7527 to 0.7632, and the 95% Hausdorff distance changes from 11.8847 to 11.0762. The results show that the proposed multi-center federated framework effectively balances privacy, security, and efficiency, significantly improves cross-center segmentation performance, and provides a new approach for 6G-enabled intelligent stroke management.

The proliferation of unprotected Internet of Things (IoT) devices has been exponential in recent years, and it will continue to rise in the years to come owing to improvements in wireless connectivity. Due to its vulnerability to malware, reliable techniques for detecting IoT malware have become imperative. Problems with non-independently and identically distributed data and poor generalizability nevertheless prevent us from reaching our objective. A methodical strategy for detecting malware is laid out in this study. Federated learning (FL) methods generate a notable degree of communication overhead given the high volumes of weights sent and received from the client-side trained models. By combining the benefits of FL with Artificial Plant Optimization Algorithm (APO), this study intends to solve this problem. APO facilitated FL framework have been assessed applying it to benchmark malware datasets. In terms of effectiveness, reliability, scalability, generalizability, and communication efficiency, the APO facilitated FL has been experimentally evaluated on readily accessible malware datasets.

With the growing demand for intelligent and precise physical education (PE) evaluation, traditional single-source data or manual scoring approaches have shown clear limitations in terms of objectivity, efficiency, and scalability. This paper proposes a novel Multimodal Attention-based Transformer-Enhanced Deep Fusion Model (MAT-DFM) for intelligent assessment in PE, leveraging wearable sensor data, video recordings, and audio signals to construct a robust, real-time evaluation framework. Through temporal synchronization, deep neural feature extraction, and attention-driven fusion, the model captures both physical performance and contextual behavioral cues. Extensive experiments on a custom multimodal PE dataset demonstrate that MAT-DFM achieves superior accuracy (91.3%) and lower MAE (3.42) compared to multiple state-of-the-art baselines, validating the effectiveness of transformer-based multimodal fusion. Furthermore, the model supports real-time feedback and fine-grained skill analysis, providing a comprehensive and scalable solution for smart PE instruction. This work presents an innovative fusion strategy that advances the development of wearable multimodal assessment systems in education.

The integration of photovoltaic (PV) smart grids with 6G networks advances sustainable energy systems but raises vulnerability to sophisticated cyber threats. Existing intrusion detection systems (IDS) suffer from static feature engineering, poor adaptability to dynamic 6G/PV environments, and high false positives, highlighting the need for a dynamically adaptive IDS. To address this, we propose an adaptive feature learning-based IDS (AFL-IDS) with three innovations: (i) mutual information-driven cross-domain fusion of PV operational and 6G traffic features; (ii) attention-guided adaptive module for real-time feature refinement; (iii) CNN–LSTM hybrid model with elastic weight consolidation to capture spatio-temporal dependencies and avoid catastrophic forgetting. Evaluated on a 150 000-instance dataset, AFL-IDS outperforms state-of-the-art baselines in key metrics. This confirms AFL-IDS as a scalable, robust framework for mitigating cyber risks in 6G-PV grids and laying a foundation for secure next-generation smart energy systems.

Against the backdrop of the rapid development of the Industrial Internet of Things (IIoT), pipeline infrastructure intelligent monitoring has become a key requirement to ensure safety and efficiency of industrial production. According to the micro-vibration signals received from precision instruments during operation or their surrounding environments, a micro-vibration detection system based on LabVIEW and Proteus was developed in this work, which integrated IIoT and achieved real-time data collection and transmission through wireless sensor networks. This research adopted the CA-YD-103 sensor as the detection device. After data collection using the NI USB-6009 data acquisition card, the micro-pressure input signal was filtered and amplified. Finally, an analysis program using the Hilbert-Huang Transform (HHT) algorithm was developed on the LabVIEW platform to draw signal time-domain and frequency-domain graphs and analyze and save the data. Micro-vibration signal and conditioning circuit simulations were performed on the Proteus platform and Multisim platforms, respectively. The results demonstrated that the system exhibited good dynamic and static properties, high accuracy, low load effect, and high anti-interference features and utilized the IIoT platform for data analysis and remote monitoring. This showed particularly high detection performance for micro-vibration micro-pressure signals (below 20 Hz and 5 V) and provided reliable technical support for intelligent manufacturing in the Industry 5.0 environment.

With the rapid evolution of sixth-generation (6G) edge networks, the highly dynamic and complex nature of network topologies poses significant challenges for threat detection. To address this issue, we propose a novel graph neural network (GNN)-based threat detection scheme (DyGNN-TD) specifically designed for dynamic topologies. The scheme introduces three key innovations: (i) a sliding-window temporal graph construction that efficiently captures short-term topological variations; (ii) a temporal attention mechanism that adaptively emphasizes recent interactions to enhance spatio-temporal modeling; and (iii) an adaptive model updating strategy that maintains robustness under frequent node churn and topological reconfigurations. By representing the 6G edge network as a temporal graph where nodes and edges encode entities and relationships, DyGNN-TD effectively learns the evolution of normal and malicious behaviors. Experimental evaluations on simulated 6G edge datasets demonstrate that DyGNN-TD achieves higher performance with baselines. These results highlight the potential of DyGNN-TD as a scalable and reliable safeguard for future 6G edge communication systems.