Pub Date : 2026-01-19DOI: 10.1109/ACCESS.2026.3655591
Mayumi Nakano;Yuya Seki;Shuta Kikuchi;Shu Tanaka
Derivative-free (DF) optimization problems aim to identify an input that maximizes or minimizes the output of an objective function whose input-output relationship is unknown. Factorization machine with quadratic-optimization annealing (FMQA) is a promising approach to this task, employing a factorization machine (FM) as a surrogate model to iteratively guide the solution search via an Ising machine. Although FMQA has demonstrated strong optimization performance across various applications, its performance often stagnates as the number of optimization iterations increases. One contributing factor to this stagnation is the growing number of data points in the dataset used to train FM. As more data are accumulated, the contribution of newly added data points tends to become diluted within the entire dataset. Based on this observation, we hypothesize that such dilution reduces the impact of new data on improving the prediction accuracy of FM. To address this issue, we propose a novel method named sliding window for iterative factorization training combined with FMQA (SWIFT-FMQA). This method improves upon FMQA by utilizing a sliding-window strategy to sequentially construct a dataset that retains at most a specified number of the most recently added data points. SWIFT-FMQA is designed to enhance the influence of newly added data points on the surrogate model. Numerical experiments demonstrate that SWIFT-FMQA obtains lower-cost solutions with fewer objective function evaluations compared to FMQA.
{"title":"SWIFT-FMQA: Enhancing Factorization Machine With Quadratic-Optimization Annealing via Sliding Window","authors":"Mayumi Nakano;Yuya Seki;Shuta Kikuchi;Shu Tanaka","doi":"10.1109/ACCESS.2026.3655591","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3655591","url":null,"abstract":"Derivative-free (DF) optimization problems aim to identify an input that maximizes or minimizes the output of an objective function whose input-output relationship is unknown. Factorization machine with quadratic-optimization annealing (FMQA) is a promising approach to this task, employing a factorization machine (FM) as a surrogate model to iteratively guide the solution search via an Ising machine. Although FMQA has demonstrated strong optimization performance across various applications, its performance often stagnates as the number of optimization iterations increases. One contributing factor to this stagnation is the growing number of data points in the dataset used to train FM. As more data are accumulated, the contribution of newly added data points tends to become diluted within the entire dataset. Based on this observation, we hypothesize that such dilution reduces the impact of new data on improving the prediction accuracy of FM. To address this issue, we propose a novel method named sliding window for iterative factorization training combined with FMQA (SWIFT-FMQA). This method improves upon FMQA by utilizing a sliding-window strategy to sequentially construct a dataset that retains at most a specified number of the most recently added data points. SWIFT-FMQA is designed to enhance the influence of newly added data points on the surrogate model. Numerical experiments demonstrate that SWIFT-FMQA obtains lower-cost solutions with fewer objective function evaluations compared to FMQA.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10977-10990"},"PeriodicalIF":3.6,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11357563","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026488","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-19DOI: 10.1109/ACCESS.2026.3655701
Savitha G. Kini;J. Lokesh;Anjan N. Padmasali
In the current era, LED lighting technology is the most widely used illumination source in all applications worldwide. Accurately predicting lumen degradation and lifetime performance has become critical for ensuring long-term reliability and cost-effectiveness. Traditional models often fail to capture the complex, non-linear nature of real-world degradation behavior. The work systematically models the lumen degradation behavior of LED luminaires using a four-parameter double exponential Gompertz function. The proposed model effectively captures the asymmetric, mirrored S-curve behavior observed in long-term degradation profiles of LED luminaires, which traditional exponential models fail to represent accurately. Experimental data from accelerated degradation tests conducted on three different commercial 16W LED luminaires were used to develop the model. The SEM-EDS analysis identified silver mirror tarnishing as a dominant physical degradation mechanism, providing material-level insight into the observed steep lumen drop during mid-life operation. A key contribution of this work is the development of a predictive framework that correlates proposed model coefficients with temperature using only three accelerated degradation tests. This enables accurate estimation of lumen maintenance performance at untested operating conditions, significantly reducing the need for exhaustive physical testing. The proposed methodology provides a practical, scalable, and cost-effective solution for predicting LED lifetime, making it highly applicable to both research and industry. It supports sustainable lighting development by improving lifetime prediction accuracy while reducing experimental burden, thereby contributing to energy-efficient operation and responsible resource utilization.
{"title":"Development of a Realistic Model to Accurately Predict the “Mirrored S-Curve” Nature of LED Luminaire Lumen Maintenance for Any Operating Conditions","authors":"Savitha G. Kini;J. Lokesh;Anjan N. Padmasali","doi":"10.1109/ACCESS.2026.3655701","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3655701","url":null,"abstract":"In the current era, LED lighting technology is the most widely used illumination source in all applications worldwide. Accurately predicting lumen degradation and lifetime performance has become critical for ensuring long-term reliability and cost-effectiveness. Traditional models often fail to capture the complex, non-linear nature of real-world degradation behavior. The work systematically models the lumen degradation behavior of LED luminaires using a four-parameter double exponential Gompertz function. The proposed model effectively captures the asymmetric, mirrored S-curve behavior observed in long-term degradation profiles of LED luminaires, which traditional exponential models fail to represent accurately. Experimental data from accelerated degradation tests conducted on three different commercial 16W LED luminaires were used to develop the model. The SEM-EDS analysis identified silver mirror tarnishing as a dominant physical degradation mechanism, providing material-level insight into the observed steep lumen drop during mid-life operation. A key contribution of this work is the development of a predictive framework that correlates proposed model coefficients with temperature using only three accelerated degradation tests. This enables accurate estimation of lumen maintenance performance at untested operating conditions, significantly reducing the need for exhaustive physical testing. The proposed methodology provides a practical, scalable, and cost-effective solution for predicting LED lifetime, making it highly applicable to both research and industry. It supports sustainable lighting development by improving lifetime prediction accuracy while reducing experimental burden, thereby contributing to energy-efficient operation and responsible resource utilization.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10860-10870"},"PeriodicalIF":3.6,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11358875","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026416","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-19DOI: 10.1109/ACCESS.2026.3655826
Tingting Guo;Sainan Yang;Yao Fu;Daitao Wang
Multi-task joint learning for complex scene image understanding faces multiple challenges, including diverse visual elements, task-specific demands, and constrained computational resources. These challenges are particularly prominent in specialized domains such as Intangible Cultural Heritage (ICH), where current research lacks effective joint modeling approaches for image classification, semantic segmentation, and object localization tasks. To address this gap, we introduce a novel multi-task visual understanding problem tailored for ICH scenarios, and construct a high-quality dataset—ICH-Scene3800—comprising 3,800 annotated images across 12 representative ICH categories. To tackle this task, we propose the first lightweight multi-task learning framework capable of performing image-level classification, instance-level localization, and instance-level detection simultaneously. The framework employs a shared backbone to learn general-purpose features and integrates an attention-guided dynamic fusion mechanism that facilitates cross-task semantic interaction. Furthermore, a group-convolution-based lightweight architecture is introduced to enable efficient feature extraction and resource-aware deployment. These designs significantly enhance the model’s generalization ability across tasks and scenes. Extensive experiments on ICH-Scene3800 and the Cityscapes dataset demonstrate that our model achieves 92.19% mIoU and 82.36% mIoU, respectively, with only 0.024M parameters and 0.085 GFLOPs. It reaches a real-time processing speed of 98.5 FPS on an NVIDIA GeForce GTX 1060 (6GB) and significantly outperforms existing methods on the LSES metric, achieving state-of-the-art performance. This research provides a practical and efficient solution for intelligent visual understanding in cultural heritage preservation and other resource-constrained application scenarios. The code and related materials are available at https://github.com/Upno111/ICH
{"title":"A Unified Lightweight Network for Complex Scene Image Understanding via Multi-Task Joint Learning","authors":"Tingting Guo;Sainan Yang;Yao Fu;Daitao Wang","doi":"10.1109/ACCESS.2026.3655826","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3655826","url":null,"abstract":"Multi-task joint learning for complex scene image understanding faces multiple challenges, including diverse visual elements, task-specific demands, and constrained computational resources. These challenges are particularly prominent in specialized domains such as Intangible Cultural Heritage (ICH), where current research lacks effective joint modeling approaches for image classification, semantic segmentation, and object localization tasks. To address this gap, we introduce a novel multi-task visual understanding problem tailored for ICH scenarios, and construct a high-quality dataset—ICH-Scene3800—comprising 3,800 annotated images across 12 representative ICH categories. To tackle this task, we propose the first lightweight multi-task learning framework capable of performing image-level classification, instance-level localization, and instance-level detection simultaneously. The framework employs a shared backbone to learn general-purpose features and integrates an attention-guided dynamic fusion mechanism that facilitates cross-task semantic interaction. Furthermore, a group-convolution-based lightweight architecture is introduced to enable efficient feature extraction and resource-aware deployment. These designs significantly enhance the model’s generalization ability across tasks and scenes. Extensive experiments on ICH-Scene3800 and the Cityscapes dataset demonstrate that our model achieves 92.19% mIoU and 82.36% mIoU, respectively, with only 0.024M parameters and 0.085 GFLOPs. It reaches a real-time processing speed of 98.5 FPS on an NVIDIA GeForce GTX 1060 (6GB) and significantly outperforms existing methods on the LSES metric, achieving state-of-the-art performance. This research provides a practical and efficient solution for intelligent visual understanding in cultural heritage preservation and other resource-constrained application scenarios. The code and related materials are available at <uri>https://github.com/Upno111/ICH</uri>","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"14916-14930"},"PeriodicalIF":3.6,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11358991","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081969","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-16DOI: 10.1109/ACCESS.2026.3654856
Mehnaz Hossain Antora;Fyaz Nafin Rahman;Rahul Debnath;Ahmed Abdelmoamen Ahmed;Md Jobair Bin Alam
Landslides pose significant risks to infrastructure, the environment, and human life, particularly when triggered by intense or prolonged rainfall. Accurate and timely prediction of slope failures remains challenging due to the complex interactions between hydrology and geotechnics, as well as spatial variability in soil conditions. This paper presents a hydrology-informed early-warning system that integrates Machine Learning (ML) and Internet of Things (IoT) technologies for real-time landslide monitoring and short-term forecasting. The proposed system is demonstrated using a laboratory-scale slope model designed as a modular monitoring unit that can be replicated across multiple locations for regional deployment. The system is trained and validated using both real-time sensor data collected from controlled laboratory rainfall experiments and synthetic hydrological datasets generated through artificial rainfall simulations that emulate large-scale, failure-prone slope conditions. Multiple ML models were developed using slope tilt measurements, soil matric suction, and soil moisture data to predict slope instability. A web-based graphical user interface enables remote, real-time visualization of sensor data. It provides automated alerts while supporting forecasts of slope failure one, two, and four days in advance. Experimental results indicate that the proposed system effectively detects slope movements and forecasts failures with acceptable accuracy, low computational overhead (approximately 148ms per sensor transmission cycle), and near real-time end-to-end latency from data acquisition to visualization. Among the evaluated models, the random forest model consistently achieved superior performance. This work establishes a validated proof of concept under controlled laboratory conditions, with future field validation on natural slopes identified as a critical next step. In addition, a cost analysis for a prospective large-scale deployment in the Greater Houston Area demonstrates the economic feasibility of the proposed monitoring architecture.
{"title":"A Scalable Hydrology-Informed Monitoring System for Early Detection of Slope Failures Using IoT and ML","authors":"Mehnaz Hossain Antora;Fyaz Nafin Rahman;Rahul Debnath;Ahmed Abdelmoamen Ahmed;Md Jobair Bin Alam","doi":"10.1109/ACCESS.2026.3654856","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3654856","url":null,"abstract":"Landslides pose significant risks to infrastructure, the environment, and human life, particularly when triggered by intense or prolonged rainfall. Accurate and timely prediction of slope failures remains challenging due to the complex interactions between hydrology and geotechnics, as well as spatial variability in soil conditions. This paper presents a hydrology-informed early-warning system that integrates Machine Learning (ML) and Internet of Things (IoT) technologies for real-time landslide monitoring and short-term forecasting. The proposed system is demonstrated using a laboratory-scale slope model designed as a modular monitoring unit that can be replicated across multiple locations for regional deployment. The system is trained and validated using both real-time sensor data collected from controlled laboratory rainfall experiments and synthetic hydrological datasets generated through artificial rainfall simulations that emulate large-scale, failure-prone slope conditions. Multiple ML models were developed using slope tilt measurements, soil matric suction, and soil moisture data to predict slope instability. A web-based graphical user interface enables remote, real-time visualization of sensor data. It provides automated alerts while supporting forecasts of slope failure one, two, and four days in advance. Experimental results indicate that the proposed system effectively detects slope movements and forecasts failures with acceptable accuracy, low computational overhead (approximately 148ms per sensor transmission cycle), and near real-time end-to-end latency from data acquisition to visualization. Among the evaluated models, the random forest model consistently achieved superior performance. This work establishes a validated proof of concept under controlled laboratory conditions, with future field validation on natural slopes identified as a critical next step. In addition, a cost analysis for a prospective large-scale deployment in the Greater Houston Area demonstrates the economic feasibility of the proposed monitoring architecture.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10735-10759"},"PeriodicalIF":3.6,"publicationDate":"2026-01-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11354470","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026414","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-15DOI: 10.1109/ACCESS.2026.3654521
Thi-Thu-Huong Le;Andro Aprila Adiputra;Anak Agung Ngurah Dharmawangsa;Hyunjin Jang;Howon Kim
The Controller Area Network (CAN) bus plays a key role in keeping vehicles safe by enabling critical systems to communicate with each other. However, because it does not have its own security features, the CAN bus is open to cyber threats. A CAN bus intrusion detection system (IDS) is critical for automotive cybersecurity. This has made it especially important to create IDS that are not just accurate but also efficient enough to run on the limited hardware of Electronic Control Units (ECUs). Unfortunately, many current deep learning solutions for CAN intrusion detection use large and complex models that are too demanding for most automotive systems. Moreover, existing deep learning approaches need excessive computational resources that are unsuitable for resource-constrained ECUs. We propose TinyCNNCANNet, an ultra-lightweight convolutional neural network with just 13K parameters, designed to provide low-latency and resource-efficient CAN intrusion detection under experimental settings. Rather than focusing on on-vehicle deployment, this work evaluates the feasibility of lightweight CNN architectures for future real-time capable CAN intrusion detection. We comprehensively evaluate TinyCNNCANNet on four diverse datasets: CANFD 2021, CICIoV 2024, Multi-Fuzzer-CAN 2025, and SynCAN 2025. These datasets encompass nine attack types. TinyCNNCANNet achieves competitive or superior performance compared to models with 115-$300times $ more parameters. All architectures detect volume-based attacks (DoS, flooding, and fuzzing) most effectively. Sophisticated attacks (malfunction and fuzzer variants) challenge all models to a similar degree, regardless of complexity. TinyCNNCANNet shows superior generalization on synthetic out-of-distribution data (SynCAN 2025). It achieves 100% accuracy, while EfficientCANNet (86.82%) and MobileNetCANNet (59.33%) fail, revealing overfitting vulnerabilities in complex models. TinyCNNCANNet delivers 12-$20times $ faster inference (0.16-0.51 ms vs. 2.14-4.15 ms) and a 145-$383times $ smaller model size (0.04 MB vs. 5.81-15.32 MB). These results demonstrate the potential of TinyCNNCANNet for real-time capable CAN intrusion detection and indicate its suitability for future deployment on embedded automotive platforms.
控制器区域网络（CAN）总线通过使关键系统能够相互通信，在保证车辆安全方面发挥着关键作用。然而，由于CAN总线没有自己的安全特性，它很容易受到网络威胁。CAN总线入侵检测系统（IDS）对于汽车网络安全至关重要。这使得创建不仅准确而且足够高效的IDS在有限的电子控制单元（ecu）硬件上运行变得尤为重要。不幸的是，目前许多用于CAN入侵检测的深度学习解决方案都使用了大型复杂的模型，这对于大多数汽车系统来说要求太高。此外，现有的深度学习方法需要过多的计算资源，不适合资源受限的ecu。我们提出了一个只有13K个参数的超轻量级卷积神经网络TinyCNNCANNet，用于在实验设置下提供低延迟和资源高效的CAN入侵检测。这项工作不是专注于车载部署，而是评估轻量级CNN架构的可行性，以实现未来实时CAN入侵检测。我们在四个不同的数据集上对TinyCNNCANNet进行了综合评估：CANFD 2021、CICIoV 2024、Multi-Fuzzer-CAN 2025和SynCAN 2025。这些数据集包含九种攻击类型。与参数多115- 300倍的模型相比，TinyCNNCANNet实现了具有竞争力或优越的性能。所有架构都能最有效地检测基于卷的攻击（DoS、泛洪攻击和模糊攻击）。复杂的攻击（故障和fuzzer变体）对所有模型的挑战程度相似，无论其复杂性如何。TinyCNNCANNet在合成分布外数据（SynCAN 2025）上表现出优越的泛化能力。它达到了100%的准确率，而效率cannet（86.82%）和MobileNetCANNet（59.33%）失败，揭示了复杂模型的过拟合漏洞。TinyCNNCANNet的推理速度提高了12- 20倍（0.16-0.51 ms vs. 2.14-4.15 ms），模型尺寸缩小了145- 383倍（0.04 MB vs. 5.81-15.32 MB）。这些结果证明了TinyCNNCANNet在实时CAN入侵检测方面的潜力，并表明其未来在嵌入式汽车平台上部署的适用性。
{"title":"Lightweight CNN-Based Intrusion Detection for CAN Bus Networks","authors":"Thi-Thu-Huong Le;Andro Aprila Adiputra;Anak Agung Ngurah Dharmawangsa;Hyunjin Jang;Howon Kim","doi":"10.1109/ACCESS.2026.3654521","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3654521","url":null,"abstract":"The Controller Area Network (CAN) bus plays a key role in keeping vehicles safe by enabling critical systems to communicate with each other. However, because it does not have its own security features, the CAN bus is open to cyber threats. A CAN bus intrusion detection system (IDS) is critical for automotive cybersecurity. This has made it especially important to create IDS that are not just accurate but also efficient enough to run on the limited hardware of Electronic Control Units (ECUs). Unfortunately, many current deep learning solutions for CAN intrusion detection use large and complex models that are too demanding for most automotive systems. Moreover, existing deep learning approaches need excessive computational resources that are unsuitable for resource-constrained ECUs. We propose TinyCNNCANNet, an ultra-lightweight convolutional neural network with just 13K parameters, designed to provide low-latency and resource-efficient CAN intrusion detection under experimental settings. Rather than focusing on on-vehicle deployment, this work evaluates the feasibility of lightweight CNN architectures for future real-time capable CAN intrusion detection. We comprehensively evaluate TinyCNNCANNet on four diverse datasets: CANFD 2021, CICIoV 2024, Multi-Fuzzer-CAN 2025, and SynCAN 2025. These datasets encompass nine attack types. TinyCNNCANNet achieves competitive or superior performance compared to models with 115-<inline-formula> <tex-math>$300times $ </tex-math></inline-formula> more parameters. All architectures detect volume-based attacks (DoS, flooding, and fuzzing) most effectively. Sophisticated attacks (malfunction and fuzzer variants) challenge all models to a similar degree, regardless of complexity. TinyCNNCANNet shows superior generalization on synthetic out-of-distribution data (SynCAN 2025). It achieves 100% accuracy, while EfficientCANNet (86.82%) and MobileNetCANNet (59.33%) fail, revealing overfitting vulnerabilities in complex models. TinyCNNCANNet delivers 12-<inline-formula> <tex-math>$20times $ </tex-math></inline-formula> faster inference (0.16-0.51 ms vs. 2.14-4.15 ms) and a 145-<inline-formula> <tex-math>$383times $ </tex-math></inline-formula> smaller model size (0.04 MB vs. 5.81-15.32 MB). These results demonstrate the potential of TinyCNNCANNet for real-time capable CAN intrusion detection and indicate its suitability for future deployment on embedded automotive platforms.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"14870-14891"},"PeriodicalIF":3.6,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11355494","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146082001","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-15DOI: 10.1109/ACCESS.2026.3654644
David Yoon Suk Kang;Eujeanne Kim;Kyungsik Han;Sang-Wook Kim
Hypergraph representation learning has gained increasing attention for modeling higher-order relationships beyond pairwise interactions. Among existing approaches, clique expansion-based (CE-based) and star expansion-based (SE-based) methods are two dominant paradigms, yet their fundamental limitations remain underexplored. In this paper, we analyze CE- and SE-based methods and identify two complementary issues: CE-based methods suffer from over-agglomeration, where node representations in overlapping hyperedges become excessively clustered, while SE-based methods exhibit under-agglomeration, failing to sufficiently aggregate nodes within the same hyperedge. To address these issues, we propose $textsf {STARGCN}$ , a hypergraph representation learning framework that constructs a bipartite graph via star expansion and employs a graph convolutional network with a tuplewise loss to explicitly enforce appropriate aggregation and separation of node representations. Experiments on seven real-world hypergraph datasets demonstrate that $textsf {STARGCN}$ consistently and significantly outperforms five state-of-the-art CE- and SE-based methods across all datasets, achieving performance gains of up to 13.2% in accuracy and 10.2% in F1-score over the strongest baseline.
{"title":"Revisiting Clique and Star Expansions in Hypergraph Representation Learning: Observations, Problems, and Solutions","authors":"David Yoon Suk Kang;Eujeanne Kim;Kyungsik Han;Sang-Wook Kim","doi":"10.1109/ACCESS.2026.3654644","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3654644","url":null,"abstract":"Hypergraph representation learning has gained increasing attention for modeling higher-order relationships beyond pairwise interactions. Among existing approaches, clique expansion-based (CE-based) and star expansion-based (SE-based) methods are two dominant paradigms, yet their fundamental limitations remain underexplored. In this paper, we analyze CE- and SE-based methods and identify two complementary issues: CE-based methods suffer from over-agglomeration, where node representations in overlapping hyperedges become excessively clustered, while SE-based methods exhibit under-agglomeration, failing to sufficiently aggregate nodes within the same hyperedge. To address these issues, we propose <inline-formula> <tex-math>$textsf {STARGCN}$ </tex-math></inline-formula>, a hypergraph representation learning framework that constructs a bipartite graph via star expansion and employs a graph convolutional network with a tuplewise loss to explicitly enforce appropriate aggregation and separation of node representations. Experiments on seven real-world hypergraph datasets demonstrate that <inline-formula> <tex-math>$textsf {STARGCN}$ </tex-math></inline-formula> consistently and significantly outperforms five state-of-the-art CE- and SE-based methods across all datasets, achieving performance gains of up to 13.2% in accuracy and 10.2% in F1-score over the strongest baseline.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10797-10810"},"PeriodicalIF":3.6,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11354166","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026373","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Neural Architecture Search (NAS) has emerged as a powerful paradigm for automating model design, yet most existing approaches remain training-intensive and computationally prohibitive. In resource-constrained domains such as UAV-based perception and Tiny Machine Learning (TinyML), performing repeated training or fine-tuning during search is infeasible due to strict compute, memory, and energy limitations. We propose a Proxy-Guided Bayesian Optimization NAS framework that eliminates all training during search by modeling a fused set of trainability proxies (e.g., SynFlow, Jacobian covariance, Neural Tangent Kernel) and hardware proxies (e.g., FLOPs, parameters, latency) within a unified Bayesian surrogate. This surrogate enables uncertainty-aware exploration directly under device-level constraints, guiding the search toward architectures that are both efficient and deployable. Unlike conventional NAS pipelines that demand extensive GPU-time for accuracy evaluations, our method completes the entire search on NATS-Bench (TSS) in only ~0.8 GPU-hours—achieving a top-1 accuracy of 93.25% with 2.10M parameters, 110M FLOPs, and 0.80 ms latency. This corresponds to an order-of-magnitude reduction in search cost compared to accuracy-driven baselines such as REA and BOHB, while preserving accuracy and satisfying all TinyML deployment budgets ($P_{max }$ , $F_{max }$ , $L_{max }$ ). By coupling hardware-awareness with training-free optimization, the proposed approach bridges the gap between proxy-based NAS and real-world, energy-efficient deployment for UAV and edge intelligence applications.
{"title":"Training-Free Proxy-Guided Bayesian NAS for UAV-Constrained TinyML","authors":"Parthiva Yadlapalli;Rishi Raj;Dayananda Pruthviraja","doi":"10.1109/ACCESS.2026.3654275","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3654275","url":null,"abstract":"Neural Architecture Search (NAS) has emerged as a powerful paradigm for automating model design, yet most existing approaches remain training-intensive and computationally prohibitive. In resource-constrained domains such as UAV-based perception and Tiny Machine Learning (TinyML), performing repeated training or fine-tuning during search is infeasible due to strict compute, memory, and energy limitations. We propose a Proxy-Guided Bayesian Optimization NAS framework that eliminates all training during search by modeling a fused set of trainability proxies (e.g., SynFlow, Jacobian covariance, Neural Tangent Kernel) and hardware proxies (e.g., FLOPs, parameters, latency) within a unified Bayesian surrogate. This surrogate enables uncertainty-aware exploration directly under device-level constraints, guiding the search toward architectures that are both efficient and deployable. Unlike conventional NAS pipelines that demand extensive GPU-time for accuracy evaluations, our method completes the entire search on NATS-Bench (TSS) in only ~0.8 GPU-hours—achieving a top-1 accuracy of 93.25% with 2.10M parameters, 110M FLOPs, and 0.80 ms latency. This corresponds to an order-of-magnitude reduction in search cost compared to accuracy-driven baselines such as REA and BOHB, while preserving accuracy and satisfying all TinyML deployment budgets (<inline-formula> <tex-math>$P_{max }$ </tex-math></inline-formula>, <inline-formula> <tex-math>$F_{max }$ </tex-math></inline-formula>, <inline-formula> <tex-math>$L_{max }$ </tex-math></inline-formula>). By coupling hardware-awareness with training-free optimization, the proposed approach bridges the gap between proxy-based NAS and real-world, energy-efficient deployment for UAV and edge intelligence applications.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10654-10666"},"PeriodicalIF":3.6,"publicationDate":"2026-01-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11352858","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026375","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-13DOI: 10.1109/ACCESS.2026.3654007
Shadi Banitaan;Taher El Taher;Khalid Aldamasi;Hassan Hassoun;Shoaib Ahmed
Accurate short-term occupancy forecasting is essential for smart building operations such as energy management, space utilization, safety, and facility planning. However, many existing solutions rely on dedicated sensors that increase deployment cost and operational complexity and limit scalability. This paper proposes a sensor-free occupancy forecasting framework that utilizes Wi-Fi syslog data already generated by enterprise networks. The study uses two real-world datasets derived from campus and office building Wi-Fi infrastructures and evaluates several machine learning models, including Random Forest, Decision Tree, Gradient Boosting, and a Long Short-Term Memory (LSTM) network, for multi-step forecasting at a 5-minute resolution. Experimental results show that Random Forest achieves the highest accuracy, with Coefficient of Determination ($R^{2}$ ) values of up to 0.997 and consistently low mean absolute error (MAE) and root mean squared error (RMSE), while LSTM provides competitive performance for short and medium forecasting horizons. Extended horizon experiments show that LSTM-based forecasts stay reliable for look-ahead periods of up to 60 minutes, while longer horizons show increased sensitivity to temporal variability and pattern changes. We also show that using only a small number of features is adequate to achieve high prediction accuracy, which simplifies data preparation and supports real-time deployment. The evaluation also examines cross-zone and cross-building generalization and demonstrates that short-term adaptation enables robust deployment across heterogeneous environments with limited retraining overhead. The proposed framework is integrated into an interactive dashboard to support visualization and decision-making. Overall, the results indicate that Wi-Fi syslog-based occupancy forecasting is a practical, scalable, and privacy-preserving approach for smart building management.
{"title":"Sensor-Free Occupancy Forecasting for Smart Buildings: A Wi-Fi Syslog Approach With Machine and Deep Learning","authors":"Shadi Banitaan;Taher El Taher;Khalid Aldamasi;Hassan Hassoun;Shoaib Ahmed","doi":"10.1109/ACCESS.2026.3654007","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3654007","url":null,"abstract":"Accurate short-term occupancy forecasting is essential for smart building operations such as energy management, space utilization, safety, and facility planning. However, many existing solutions rely on dedicated sensors that increase deployment cost and operational complexity and limit scalability. This paper proposes a sensor-free occupancy forecasting framework that utilizes Wi-Fi syslog data already generated by enterprise networks. The study uses two real-world datasets derived from campus and office building Wi-Fi infrastructures and evaluates several machine learning models, including Random Forest, Decision Tree, Gradient Boosting, and a Long Short-Term Memory (LSTM) network, for multi-step forecasting at a 5-minute resolution. Experimental results show that Random Forest achieves the highest accuracy, with Coefficient of Determination (<inline-formula> <tex-math>$R^{2}$ </tex-math></inline-formula>) values of up to 0.997 and consistently low mean absolute error (MAE) and root mean squared error (RMSE), while LSTM provides competitive performance for short and medium forecasting horizons. Extended horizon experiments show that LSTM-based forecasts stay reliable for look-ahead periods of up to 60 minutes, while longer horizons show increased sensitivity to temporal variability and pattern changes. We also show that using only a small number of features is adequate to achieve high prediction accuracy, which simplifies data preparation and supports real-time deployment. The evaluation also examines cross-zone and cross-building generalization and demonstrates that short-term adaptation enables robust deployment across heterogeneous environments with limited retraining overhead. The proposed framework is integrated into an interactive dashboard to support visualization and decision-making. Overall, the results indicate that Wi-Fi syslog-based occupancy forecasting is a practical, scalable, and privacy-preserving approach for smart building management.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10891-10909"},"PeriodicalIF":3.6,"publicationDate":"2026-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11348122","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026376","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Practical and intuitive communication remains a critical challenge in Human-Robot Collaboration, particularly within domestic environments. Conventional systems typically rely on structured (scripted) speech inputs, which may limit natural interaction and accessibility. This study evaluates user preferences and system usability between structured and unstructured (conversational) speech modalities in a collaborative cooking scenario using a mobile manipulator robot. Thirty adult participants engaged in tasks involving both communication modes, during which the frequency and impact of robot execution errors were also assessed. The proposed Speech2Action framework integrates Google Cloud Speech-to-Text, BERT, and GPT-Neo models for intent recognition and command generation, combined with ROS-based motion control for object retrieval. Usability and perception were analyzed using System Usability Scale (SUS) and Human–Robot Collaboration Questionnaire (HRCQ) metrics through paired t-tests and correlation analyses. Results show a preference for unstructured speech (p = 0.0032) with higher SUS scores, while robot execution errors affected perceived safety but not overall usability, consistent with the Pratfall Effect. The findings inform the design of natural, robust, and user-centric speech interfaces for collaborative robots.
{"title":"Structured and Unstructured Speech2Action Frameworks for Human–Robot Collaboration: A User Study","authors":"Krishna Kodur;Manizheh Zand;Matthew Tognotti;Cinthya Járegui;Maria Kyrarini","doi":"10.1109/ACCESS.2026.3653715","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3653715","url":null,"abstract":"Practical and intuitive communication remains a critical challenge in Human-Robot Collaboration, particularly within domestic environments. Conventional systems typically rely on structured (scripted) speech inputs, which may limit natural interaction and accessibility. This study evaluates user preferences and system usability between structured and unstructured (conversational) speech modalities in a collaborative cooking scenario using a mobile manipulator robot. Thirty adult participants engaged in tasks involving both communication modes, during which the frequency and impact of robot execution errors were also assessed. The proposed Speech2Action framework integrates Google Cloud Speech-to-Text, BERT, and GPT-Neo models for intent recognition and command generation, combined with ROS-based motion control for object retrieval. Usability and perception were analyzed using System Usability Scale (SUS) and Human–Robot Collaboration Questionnaire (HRCQ) metrics through paired t-tests and correlation analyses. Results show a preference for unstructured speech (p = 0.0032) with higher SUS scores, while robot execution errors affected perceived safety but not overall usability, consistent with the Pratfall Effect. The findings inform the design of natural, robust, and user-centric speech interfaces for collaborative robots.","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"10782-10796"},"PeriodicalIF":3.6,"publicationDate":"2026-01-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11348049","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026288","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/ACCESS.2026.3651866
Jung Yeon Hwang;Jong Hwan Park
Asymmetric broadcast encryption (ABE) allows a sender, given the public keys or identities of recipients, to encrypt a message such that only an authorized subset of users can decrypt it. In fully asymmetric settings, where any user may act as a sender, ciphertext generation time and ciphertext size become critical performance metrics. However, most existing ABE schemes impose substantial sender-side computational costs and scale poorly with system size. This paper presents new ABE constructions that achieve fast ciphertext generation while maintaining compact ciphertexts. Our schemes are built upon the identity-based revocation (IBR) framework, enabling each user’s identity to serve directly as a public key. We first propose a basic IBR scheme that produces constant-size ciphertexts independent of the number of recipients or revoked users, achieving efficient encryption through optimized hash-to-point and aggregation techniques. We then extend this design to a tree-based construction that supports large-scale systems and offers a practical trade-off among encryption cost, decryption efficiency, and secret-key size. Both schemes are proven CPA-secure under a modified Decisional Bilinear Diffie–Hellman (mDBDH) assumption in the random-oracle model. Extensive experiments with concrete parameters demonstrate that our schemes significantly outperform existing asymmetric revocation approaches. For a system with $10^{6}$ users and a revocation rate of 1.5–3%, prior schemes require tens of seconds to generate a ciphertext, whereas our constructions complete encryption within 1.6 seconds while keeping the ciphertext size nearly constant (below $10^{2}$ KB).
{"title":"Constructing Identity-Based Revocation Schemes for Efficient Generation of Ciphertexts","authors":"Jung Yeon Hwang;Jong Hwan Park","doi":"10.1109/ACCESS.2026.3651866","DOIUrl":"https://doi.org/10.1109/ACCESS.2026.3651866","url":null,"abstract":"Asymmetric broadcast encryption (ABE) allows a sender, given the public keys or identities of recipients, to encrypt a message such that only an authorized subset of users can decrypt it. In fully asymmetric settings, where any user may act as a sender, ciphertext generation time and ciphertext size become critical performance metrics. However, most existing ABE schemes impose substantial sender-side computational costs and scale poorly with system size. This paper presents new ABE constructions that achieve fast ciphertext generation while maintaining compact ciphertexts. Our schemes are built upon the identity-based revocation (IBR) framework, enabling each user’s identity to serve directly as a public key. We first propose a basic IBR scheme that produces constant-size ciphertexts independent of the number of recipients or revoked users, achieving efficient encryption through optimized hash-to-point and aggregation techniques. We then extend this design to a tree-based construction that supports large-scale systems and offers a practical trade-off among encryption cost, decryption efficiency, and secret-key size. Both schemes are proven CPA-secure under a modified Decisional Bilinear Diffie–Hellman (mDBDH) assumption in the random-oracle model. Extensive experiments with concrete parameters demonstrate that our schemes significantly outperform existing asymmetric revocation approaches. For a system with <inline-formula> <tex-math>$10^{6}$ </tex-math></inline-formula> users and a revocation rate of 1.5–3%, prior schemes require tens of seconds to generate a ciphertext, whereas our constructions complete encryption within 1.6 seconds while keeping the ciphertext size nearly constant (below <inline-formula> <tex-math>$10^{2}$ </tex-math></inline-formula> KB).","PeriodicalId":13079,"journal":{"name":"IEEE Access","volume":"14 ","pages":"7730-7743"},"PeriodicalIF":3.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11339495","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982135","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：