Semantic segmentation is a fundamental component of Internet-enabled visual sensing networks, where distributed cameras, edge devices, and networked sensors collect large-scale visual data from complex environments. However, such data often exhibit severe pixel-level imbalance and category heterogeneity, with semantic classes differing in frequency, spatial extent, boundary complexity, and structural patterns. These characteristics make standard pixel-wise training prone to biased decision boundaries and insufficient representations for underrepresented categories. In this paper, we identify two key limitations in this setting: the misalignment between standard optimization objectives and class-balanced evaluation criteria, and the lack of class-adaptive feature aggregation under heterogeneous category statistics. To address them, we propose a unified category-aware prior-compensation framework that combines output-level calibration with sampling-level adaptive feature learning. Specifically, we adopt relative-prior logit adjustment to alleviate pixel-wise decision bias toward dominant classes. Meanwhile, we introduce a spatial-demand-aware class-adaptive deformable attention mechanism that modulates deformable offsets during training according to class frequency and class-wise spatial demand. This design encourages flexible contextual aggregation for underrepresented and spatially complex categories while maintaining stable aggregation for dominant ones. During inference, the offset predictor remains fully feature-driven and requires no ground-truth labels, pseudo-labels, or external class annotations. Experiments on Cityscapes and ADE20K show that the proposed method improves segmentation performance for underrepresented classes while maintaining competitive overall accuracy, achieving superior class-balanced performance over representative baselines.
Fatigue driving is a leading cause of traffic accidents, necessitating robust and real-time detection methods. However, existing facial feature-based approaches often suffer from significant performance degradation under low-light conditions and facial occlusions. To address these challenges, we propose an enhanced RSE-YOLOv8n framework. Specifically, a Retinexformer-based module is introduced for illumination compensation, SAC is employed to enable efficient multi-scale feature extraction while reducing computational complexity, and an Efficient Group-wise Multi-scale Attention (EGMA) mechanism is integrated to refine feature representation. Furthermore, the dataset is augmented with occlusion and head pose variations to improve generalization. Experimental results demonstrate that the proposed method achieves a 4.5% improvement in mAP@0.5 over the baseline YOLOv8n, reaching 96.9%, while maintaining real-time inference capability. The proposed framework exhibits strong robustness and effectiveness for fatigue driving detection in challenging environments.
Service recommendation in edge Service-Oriented Architecture (SOA) environments must balance privacy protection, recommendation accuracy, and the limited computing resources of user devices. Many existing federated recommendation methods protect raw interaction data, but often still rely on broad client participation, complex scoring models, or user-feature uploads for client grouping. This paper proposes a Dual-Personalized Federated Fast Convergence Model (DPFedFast) for efficient service recommendation. Compared with FedFast, DPFedFast introduces three technical improvements: first, clients are sampled according to the similarity of locally trained rating-function parameters rather than explicit user information; second, the recommendation model is simplified by adopting a lightweight single-layer neural rating function to reduce edge-side computation and communication; and third, a dual personalization mechanism separately updates personalized rating-function parameters and item embeddings to preserve user-specific preferences under the lightweight design. Extensive experiments conducted on four real-world datasets demonstrate the effectiveness of DPFedFast.
Botnets represent a considerable threat to network security due to their ability to exploit compromised systems for coordinated malicious activities. While many existing botnet detection models achieve high detection performance, they often fail to address the vital aspect of computational efficiency. This research presents a new machine learning framework called MAGNET-Bot that employs a multilevel ensemble feature selection approach designed to improve detection performance while reducing the execution time. Unlike previous studies that typically rely on straightforward selection methods, the proposed approach incorporates multiple ensemble feature selection techniques integrated with machine learning classification models. To evaluate the effectiveness of this method, graph metadata extraction results from network traffic were analyzed using three recognized datasets: CTU-13 (characterizing sporadic behavior), NCC (representing periodic behavior), and NCC-2 (exhibiting simultaneous behavior). Experimental results indicate significant improvements in detection performance compared to traditional methods. Notably, the combination of the proposed feature selection method with the XGBoost algorithm achieved an average accuracy of 97.52%, a precision of 43.31%, a recall of 47.80%, while ensuring computational efficiency with a processing time of merely 0.12 s. Given its efficiency, performance, and thorough analytical approach, this proposed method is deemed suitable to assist network administrators in conducting forensic investigations of botnet attacks.
This paper focuses on the key problem of maximizing user service coverage under a limited deployment budget in edge computing scenarios. To meet the demand of service providers to select k locations from a set of candidate edge servers for model deployment so as to maximize coverage of end-user groups, we formally define the k-Model Deployment Coverage Maximization (kMD-C) problem and prove that it is NP-hard. For this problem, we propose two methods: MD-O, an exact benchmark method based on integer programming suitable for small-scale problems, and MD-A, an efficient heuristic algorithm with superior empirical performance. Experimental validation on real-world datasets shows that MD-A significantly outperforms several state-of-the-art baseline methods in terms of coverage performance and computational efficiency, demonstrating its superior practicality and scalability for solving large-scale kMD-C problems. This study provides a theoretical foundation and efficient tools for the scientific deployment of edge intelligent services, and helps improve service coverage and user satisfaction under resource-constrained conditions.
