Internet Technology Letters最新文献

Multimedia Wireless Sensor Networks (MWSNs) form a crucial backbone of smart cities, enabling real-time data collection and communication across diverse applications. However, ensuring secure routing, authentication, and efficient data transmission remains a major challenge due to limited resources and susceptibility to malicious attacks. This study proposes a trust-based routing and authentication framework that integrates lightweight cryptographic mechanisms with a three-phase certification process involving key generation, encryption, and decryption. A Trust Authority (TA) manages node registration, certificate issuance, and revocation, while cluster head (CH) selection is optimized using energy and distance-based formulations to balance load and reduce energy consumption. The model employs both identification-based and control-packet authentication, ensuring confidentiality, non-repudiation, and protection against internal attacks. Experimental evaluation on MATLAB with 100–500 sensor nodes and varying malicious ratios demonstrates that the proposed framework achieves higher detection accuracy, lower certification delay, reduced energy consumption, and improved throughput compared to existing approaches, making it suitable for secure smart city deployments.

In view of the challenges faced by traditional teaching models in the context of digital transformation of education, this study proposes to build a smart education environment based on the industrial Internet and innovate the practical teaching mode of normal students. The research adopts hierarchical system architecture to integrate data collection, edge computing and cloud computing technologies, and focuses on optimizing the support vector machine algorithm to achieve educational data classification and anomaly detection, with an accurate rate of 93.7%. Experimental results show that multimodal data fusion improves the analysis accuracy by 15%, and the real-time feedback delay is controlled within 200 ms, which effectively supports teaching evaluation and behavior analysis.

With the rapid deployment of photovoltaic (PV) systems and the transition toward decentralized energy infrastructures, traditional centralized economic dispatch methods are increasingly challenged by scalability bottlenecks, communication overhead, and vulnerability to single-point failures. These issues are further exacerbated by the dynamic and distributed nature of PV-based microgrids, where plug-and-play devices, intermittent generation, and privacy constraints demand localized decision-making and coordination. To address these challenges, this paper proposes a fully distributed economic dispatch framework based on a multi-agent system and a Modified Exact Diffusion Algorithm (MEDA). The framework models PV units, battery storage, flexible loads, and grid interfaces as autonomous agents that interact through peer-to-peer communication, collaboratively achieving global optimality without centralized supervision. A SOC-aware battery cost model, dynamic electricity pricing, and quadratic line loss modeling are integrated to enhance practical realism. Simulation results on a modified IEEE 33-bus microgrid show that the proposed approach significantly outperforms centralized and existing distributed methods in terms of cost reduction, convergence speed, resilience to communication failures, and adaptability to agent dynamics.

Deep learning and wearable devices for heart rate detection have been widely applied in sports for real-time body monitoring. However, existing deep networks such as convolutional networks (CNNs) and recurrent neural networks (RNNs) are unable to model the spatiotemporal features of time series signals. Moreover, these models are unable to model the uncertainty in complex motion scenes. To this end, this article constructs an effective abnormal heart rate detection system based on a variant variational autoencoder. First, the photoplethysmography (PPG) signals from different user terminals are collected and transmitted to the server through the wireless sensor network. Then, on the server side, we deployed a novel variant variational autoencoder (VAE) by exploiting the 1D convolution operation and the temporal convolutional network (TCN) module for spatiotemporal feature extraction of time series. Moreover, the VAE can effectively alleviate uncertainty in motion scenes. Finally, we conducted comparative experiments on our self-built dataset of abnormal heart rate during exercise, and the experimental results showed that the proposed model achieves the highest anomaly detection performance.

Real-time multi-object detection on smartphones requires a careful balance of accuracy, latency, energy efficiency, and data privacy. We introduce FedEdgeDetect, a unified framework that combines federated learning with edge-assisted inference to address these challenges holistically. The system incorporates a hardware-aware YOLOv5s variant with lightweight attention modules for efficient on-device execution. A capability-clustered federated training protocol is designed to ensure privacy through differential noise injection and secure aggregation, while reducing communication overhead. At inference time, a dynamic controller adaptively partitions computation between the device and edge, optimizing for real-time performance and energy consumption. Experiments across diverse datasets and devices demonstrate that FedEdgeDetect consistently improves detection accuracy, accelerates inference, enhances energy efficiency, and enforces strong privacy guarantees, outperforming existing mobile detection baselines.

This paper proposes a novel hybrid framework that combines lightweight edge-side intent sketching with cloud-based large language model (LLM) reasoning, called Wireless LLM-Enhanced Intent-Service Parsing Framework (WISE). Specifically, WISE integrates four components: Local Intent Sketching Module (LISM), Semantic Feature Compression and Transmission (SFCT) unit, Prompt-Aware LLM Service Classification Engine (LSCE), and Semantic Alignment and Service Prediction module (SASP). This architecture enables efficient semantic understanding with minimal transmission overhead. Experimental results on a curated English tutoring intent-service dataset demonstrate that WISE achieves superior accuracy (88.9% intent classification accuracy and 86.5% F1 score), while reducing communication costs by over 80% compared to cloud-only LLM solutions. Additional ablation studies and training analyses confirm the effectiveness and stability of the proposed design. WISE offers a scalable and real-time solution for intelligent language tutoring in wireless edge environments.

This letter presents an enhanced convolutional neural network (CNN)–transformer architecture integrated with large model (LM) capabilities for adaptive music quality classification in wireless communication networks (WCNs). The proposed approach combines the global feature learning strength of transformer encoders with the local pattern recognition abilities of CNNs while leveraging LM knowledge for improved audio signal understanding. To enhance classification accuracy, we first preprocess the music signal data through channel-aware normalization and feature standardization. Subsequently, we employ a multi-head attention mechanism from transformer networks to capture long-range dependencies in music features affected by wireless transmission while utilizing CNN layers to extract localized audio patterns. Finally, we incorporate inception modules to achieve multi-scale feature fusion and complete the music quality classification task. Experimental validation on the MusicCaps dataset demonstrates that our model achieves 97.8% classification accuracy, with precision, recall, and F1-score all exceeding 97.5%, outperforming existing approaches for music quality assessment in wireless environments.

This letter presents a novel framework combining large language models with a sparse attention vision transformer (SA-ViT) to classify English teaching video traffic in wireless networks. Our approach analyzes both visual content frames and extracted English speech transcripts to identify educational content types, difficulty levels, and priority requirements. The proposed model transforms video frames into visual patches while simultaneously processing English linguistic content through pre-trained language models, enabling an understanding of educational multimedia traffic. Through extensive evaluation of real-world English teaching video datasets transmitted over wireless networks, our SA-ViT framework achieves 97.5% classification accuracy, representing an 11.3% improvement over conventional video traffic classification methods. The results demonstrate effective integration of visual understanding, English language comprehension, and wireless network optimization for enhanced educational content delivery.

Modern physical education increasingly relies on wireless communication networks to deliver immersive training experiences through wearable devices, motion-tracking sensors, and real-time performance analytics. However, optimizing wireless network performance in dynamic physical education environments presents complex challenges due to rapidly changing user mobility patterns, varying signal interference from athletic equipment, and fluctuating bandwidth demands during different exercise activities. This letter proposes a novel wavelet-enhanced large model framework that integrates wavelet transform signal processing with enhanced position encoding in transformer architectures to predict and optimize wireless network performance for physical education applications. Experimental validation demonstrates that our proposed model accurately captures non-stationary behavior and abrupt changes in wireless network performance during various physical activities. The RMSE and MAPE metrics show improvements of 29.9% and 2.9%, respectively, compared to baseline transformer models, and 34.5% and 3.4% improvements compared to LSTM approaches, providing a novel technical solution for smart physical education network management.

Autonomous vehicles often struggle in dense urban intersections because of occlusions and reactive single-vehicle planners. SafeEdge tackles this challenge by partitioning cooperative motion planning across a three-tier mobile-edge hierarchy. A graph-transformer intention generator running on roadside and metro-edge nodes fuses V2X trajectory snippets from up to 120 agents into kilobyte-size probability maps of future maneuvers. Each vehicle keeps feasibility checks on board, while large-scale collision resolution is off-loaded to a metro-edge mixed-integer solver. A Coq-verified safety shield triggers emergency braking whenever network latency exceeds a derived bound. Deployed on 10 Jetson Orin NX cars and 4 Xeon Silver edge servers over a standalone 5 G link, SafeEdge clears four-way intersections with a 96% success rate and a 95th-percentile decision latency of 32 ms—well below the 50 ms safety envelope. Relative to an on-board MPC baseline, emergency-brake events drop by 70% and energy per kilometer falls by 11%. These results demonstrate that intention-aware, edge-partitioned planning can simultaneously satisfy real-time and safety requirements in dense urban driving.