Dynamic web-client programming with HTML, CSS, and JavaScript has become dominant in web application system. To assist its study by novices, we have developed a Phrase Fill-in-Blank Problem (PFP) in the Web-client Programming Learning Assistant System (Web-PLAS). Any answer is marked through string matching with the correct one. However, the current blank phrase selection algorithm implemented for helping a teacher to make a new PFP instance may not select blanks having unique answers. In this paper, we investigate a generative-AI approach to validating the selected blank phrases by the algorithm. In this approach, we first change the name of any user-defined identifier in a code so that the AI cannot use a memorized code for the answer. Then, we check the validity of the selected blanks by asking the AI to find their correct answers. After that, we repeat selecting a new phrase for blank and validating it by asking the AI to find the correct answer, until no new phrase can be selected. In either case, if the AI cannot find the correct answer, the blank is discarded. For preliminary evaluations, we manually apply this approach to the 10 PFP instances for basic topics in our previous study, using MS Copilot. The results show that several blanks in existing instances are not correct and additional blanks can be selected. The implementation of a program to automate the approach and its comprehensive evaluations with various codes will be in our next study.
Smart cities increasingly rely on computer vision systems to understand and respond to local urban dynamics. However, the effectiveness of such systems depends on datasets tailored to local environments and needs, which are often scarce or unavailable. Data scarcity arises from regional variability, privacy concerns, and the high cost of data collection and annotation. This paper proposes a synthetic-data generation and validation framework for rapidly developing locally adaptive perception models in smart-city environments. Using electric scooter detection as a representative case study, we demonstrate how diffusion-based generative models can create realistic, diverse, and privacy-preserving images to supplement limited real-world datasets. We evaluate three state-of-the-art diffusion models, perform a comparative analysis, and report on average 6% improvement in detection performance across these models when training YOLO-based object detectors. Results show that synthetic augmentation substantially improves detection accuracy in data-scarce scenarios, validating the utility of synthetic data for local adaptation. The resulting model is further deployed on an edge device as a proof of concept, illustrating the feasibility of lightweight, privacy-aware, and adaptable vision systems for next-generation smart cities.
The rapid growth of the Internet of Things (IoT) is driving the demand for miniaturized, energy-efficient, and low-cost devices with integrated wireless functionality. However, conventional PCB and chip antennas constrain form-factor flexibility and hinder rapid customization for space-constrained IoT platforms. This work addresses these limitations by designing, fabricating, and experimentally evaluating a compact inverted-F antenna (IFA) produced via Additive Microelectronics Manufacturing (AME). The antenna is optimized through full-wave simulations, prototyped in both isolated and integrated configurations, and validated using S-parameter and link-quality measurements, demonstrating performance comparable to commercial FR-4 PCB antennas and superior to most chip antennas.
Smart contracts (SC) are deployed on blockchain platforms without sufficient descriptive metadata. This lack of metadata significantly limits their discoverability, making it challenging for consumers to identify and select suitable contracts. This paper presents a comparative analysis of existing smart contract registry design approaches. It first identifies the core design properties of smart contract registries, such as descriptive metadata, description format, registry type, structure, platform scope, and discovery mechanism, and uses these dimensions as the foundation for our analysis. Drawing from this analysis, we propose a two-part taxonomy that distinguishes between purpose-based and business-based designs. We then analyze representative smart contract registry designs and highlight current key challenges in the field. By mapping the strengths, limitations, and trade-offs of these designs, this paper provides a reference to guide developers and platform architects. It aims to support the design of more efficient, transparent, interoperable, and decentralized smart contract registries, ultimately improving contract discoverability and selection for consumers.
This paper proposes a real-time semantic-depth system that fuses a Single-Beam LiDAR with a general-purpose camera. Cameras and LiDAR are critical sensors for environmental perception in autonomous systems, primarily being used for object recognition and distance measurement. Although cameras offer rich visual data, they lack direct distance measurement and are vulnerable to lighting changes. In contrast, LiDAR provides precise 3D geometric data, but its output is sparse and lacks visual detail. The proposed system projects the LiDAR point cloud onto the camera image, implementing semantic mapping by assigning meaningful information to each object via a semantic segmentation model. Experimental results demonstrate that the system performs both object recognition and accurate distance estimation, with computed RMSE (Root Mean Squared Error) and MAE (Mean Absolute Error) for each object class. This approach is applicable in low-cost hardware environments, making it suitable for applications in autonomous driving and robotics.
Mitigating urban heat requires high-resolution estimation at the block–building scale; however, conventional interpolation approaches depend heavily on sensor layout, and tabular encodings of features discard spatial context such as boundaries and connectivity. We propose a non-interpolative patch-regression CNN that treats land use/land cover (LULC) and shadow as multi-channel images and fuses them with a small set of external meteorological covariates to estimate near-surface air temperature. Fixed and mobile observations are used solely as supervision during training, so inference requires no local sensor observations. In an evaluation over Tohoku University Aobayama (2m resolution; time-independent inputs), the proposed model achieved MAE comparable to distance-based Random Forest (RF) and Gradient Boosting (GBR), while delivering consistently higher R2 (up to 52%). Predicted temperature maps captured plausible spatial variability aligned with shadows and land use. Moreover, ablation studies indicate that the shadow and LULC channels, as well as incorporating mobile observations during training, each contribute to improved explanatory power.
The integration of artificial intelligence (AI) into industrial environments increasingly depends on the availability of representative and physically consistent datasets. However, real manufacturing data are often proprietary, fragmented, or system-specific, limiting the development and validation of generalized learning models. To address this gap, this paper introduces a physics-informed synthetic dataset derived from first principles and calibrated using empirical ranges observed in an actual conveyor-furnace system. The proposed dataset integrates simplified thermodynamic and kinematic equations to preserve realistic relationships among temperature, residence time, and energy consumption. It enables two complementary applications: (i) AI-driven parameter recommendation for optimizing throughput and energy efficiency, and (ii) production forecasting based on temporal process variables. Experimental results using tree-based regression models demonstrate that the generated data are physically coherent and suitable for both optimization and predictive analytics. The proposed methodology contributes a reproducible benchmark for evaluating intelligent control and forecasting algorithms in smart manufacturing contexts.
Zero-day attacks pose a major risk and are an evolving threat to cybersecurity, as they exploit software vulnerabilities that evade traditional signature-based detection methods that often fail to recognize them. To address this challenge, machine learning based anomaly detection approaches have gained increased attention for their potential to identify unusual behaviors of emerging threats attack patterns proactively. Among these, two approaches Isolation Forest and Autoencoder models use unsupervised techniques for anomaly detection, to distinguish malicious activities. These are selected for their complementary strengths: Isolation Forest for efficient anomaly isolation in high-dimensional data, and Autoencoder for capturing complex nonlinear data patterns. This paper presents a comparative study of these two models to evaluate their ability to detect proactive zero-day attack detection, using a UGRansome dataset exclusively, as it specifically targets ransomware and zero-day attack behaviours in network traffic. The study aims to evaluate the performance, strengths and limitations of each model in detecting zero-day threats. Through systematic experimentation the research contributes to enhancing proactive security solutions.
Diabetic foot disease is a leading cause of preventable disability, lower-limb amputation, and mortality worldwide. Loss of sensation in the feet for individuals suffering from diabetes is common. Risk of lack of ability to perceive injuries to the feet is dangerous. Our team has developed a foot-monitoring system aiming to minimize this risk. Continuous monitoring of plantar pressure together with in-shoe temperature and humidity has been shown to reduce ulcer risk, but is bulky, and difficult to deploy at scale. At the same time, data-driven gait classification approaches increasingly rely on deep learning models whose memory and compute requirements exceed the capabilities of microcontroller-class wearable hardware. We present a low-cost, multimodal solution that combines an array of force-sensing resistors (FSRs), an inertial measurement unit (IMU), temperature and humidity sensing, and an ultrasonic distance sensor for adaptive power management. We describe a complete pipeline from hardware design, and wearable environmental test chamber construction, to enable controlled multimodal data collection. Soft probability outputs from TabPFN were distilled into the sub-kilobyte MLP.
Medical transcription remains a cornerstone of administrative AI in healthcare, yet conventional solutions often rely on cloud infrastructure that introduces latency, increases costs, and compromises data privacy. This paper introduces a fully edge-deployed medical transcription system designed to operate autonomously on local devices using Intel Core Ultra series (AI PC) platform equipped with heterogeneous compute resources (CPU, GPU and NPU). By combining lightweight AI models with hardware-aware optimization, the system delivers real-time transcription performance tailored for clinical environments. Beyond transcription, the system integrates seamlessly with Electronic Medical Record (EMR) platforms using Large Language Model (LLM) assistance, enabling structured documentation with minimal clinician intervention. Empirical results show that the edge-based approach achieves faster response times (28% better), maintains competitive accuracy (0.81 BERT score), and significantly reduces operational costs (44% over 5 years) compared to cloud-dependent alternatives. These findings demonstrate a scalable and privacy-conscious pathway for deploying AI in healthcare workflows, addressing longstanding integration challenges and paving the way for more efficient clinical documentation.
This paper presents a systematic study of stack optimization strategies for the NIST PQC standard ML-DSA signature verification algorithm in resource-constrained devices such as IoT devices, sensors, and security modules with limited RAM. Building on the optimization techniques proposed for Dilithium by Bos et al. (AFRICACRYPT 2022), we reorganize them into nine distinct strategies and systematically evaluate their cumulative impact on Cortex-M processors with hardware acceleration, using the Samsung Exynos 2500 security subsystem as a representative platform. Our analysis shows that stack optimization must balance memory savings while maintaining hardware acceleration and account for memory placement constraints, such as whether keys and signatures are stored internally or externally. This balance becomes especially critical in devices where verification time accumulates across multiple signatures. Overall, the findings provide practical guidance for diverse deployment scenarios, from extremely memory-limited devices to accelerator-enabled devices and balanced devices such as security modules.
Remote health monitoring is of great potential to bring long-term continuous big data for precision medicine. The challenge arises since the noise and interferences usually contaminate the signal significantly. In this study, to decode the medical insights from the rapidly decayed signal captured remotely, we propose a deep multi-channel signal decoding system with cascaded latent-space fidelity boosting. The system is composed of the multi-channel acoustic sensing device and the deep multi-channel signal decoding algorithm. The device can be placed on the bedframe to capture the spatial dynamics of the acoustic information during breathing. The deep learning framework with cascaded latent-space fidelity boosting before inference, can suppress noise and interferences hidden in the signal, through multiple iterations of space transformation and signal reconstruction. We take the sleep apnea as the real-world study, to demonstrate the promise of the proposed system. Our experiments have demonstrated the effectiveness of the system, superior to state-of-the-art. This study will promisingly advance remote signal perception and mining for precision medicine.
The study proposes an artificial intelligence– based system for simultaneous distraction detection, using the YOLOv8 model to identify cellphone use and MediaPipe to detect drowsiness, with audible and visual alerts that serve as an immediate response. The system was implemented in Python and evaluated in a controlled environment with everyday drivers, considering different lighting conditions and video resolutions (720p and 1080p). In the experiments, the YOLOv8m model achieved an mAP = 0.50 and 20.5 FPS on an RTX 3050 GPU, and 5 FPS on an i5 CPU. The results confirm the solution's feasibility and the real-time operation of the alerts, with the temporal analysis documented in a Gantt chart.
Adversarial attacks pose a growing challenge to modern smart-city vision systems, where misclassification in autonomous vehicles, surveillance platforms, and traffic-control sensors can have direct safety implications. To assess these vulnerabilities, we introduce X2M, a cross-norm dual-momentum adversarial attack designed to produce highly transferable and perceptually similar perturbations using an ensemble of lightweight and mid-scale surrogate models. X2M demonstrates reliable effectiveness against CNNs, Transformers, and adversarially trained networks across ImageNet, GTSRB, COCO-mini, and the real-world HIT-UAV infrared dataset. It maintains more than 99% visual similarity on COCO-mini and HIT-UAV, achieves near perfect white-box success on CNN and YOLO classifier backbones, obtains 94% white-box success on ViT-B, and reaches by 56% transfer success on Swin-T under strict black-box settings. These results highlight the importance of comprehensive adversarial robustness evaluation prior to deploying AI driven vision components in safety critical consumer and smart-city environments.
Reliable cooperative estimation in multi-agent systems becomes challenging when sensor noise and communication sparsity destabilize classical consensus, often causing drift, delayed convergence, or full breakdown in extreme regimes. We propose the Diffusive Cognitive Consensus (DCC) framework, a lightweight and modular architecture that augments consensus with conditional, structure-aware diffusion activated only when multi-criteria gates detect instability. The framework introduces amortized denoising, adaptive gain, rollback protection, and center-preserving corrections, ensuring bias-free and reversible updates with negligible computational overhead. Across 36 configurations, the method improves terminal accuracy by up to 13.9% over consensus in extreme noise, achieves significant gains in 15 regimes, and maintains ultra-conservative intervention rates (FracDiff < 3%). Compared to classical filters, it reduces final error by up to 7.4×, demonstrating the superiority of targeted, interpretable, post-consensus refinement. These results establish the approach as a practical and safety-aware mechanism for stabilizing cooperative behavior in noisy and partially connected multi-agent systems.
Real-time detection of anomalous behaviors in CCTV footage remains challenging due to the rarity of incidents, varied environments, and class imbalance. While large Vision-Language Models like CLIP have demonstrated robust image understanding, they lack temporal modeling needed for video anomaly recognition. In this work, we propose a temporal-aware CLIP-based framework to detect abnormal events in surveillance video. The model integrates a lightweight temporal transformer and a multimodal fusion mechanism that combines visual features with class-specific text embeddings, enabling context-aware video understanding. To address the scarcity of rare events, we augment the training data with synthetic video clips generated using OpenAI’s Sora text-to-video model. Experiments on a public CCTV anomaly dataset show that our enhanced CLIP model outperforms a baseline image-level CLIP classifier, with substantial improvements in precision, recall, and F1-score. In particular, synthetic augmentation boosted the detection of infrequent behaviors like "fight" and "fall down." These results demonstrate the practical benefit of combining pre-trained vision-language representations with generative data augmentation for anomaly detection. We also discuss the contributions of each component and remaining challenges.
The Student Career Pathway Analyzer (SCPA) is an AI-powered career guidance platform designed to bridge the gap between academic preparation and real-world employment. It provides personalized career recommendations by interpreting a student’s academic record, skills, and experiences to suggest relevant career trajectories across multiple disciplines. Beyond simple job matching, SCPA offers reverse skill mapping, career readiness insights, and interactive support to help students understand and address their individual skill gaps. By synthesizing academic data with real-time labor market information, the platform enables a holistic view of each student’s professional profile. The system empowers learners to make informed career decisions and equips institutions with scalable, data-driven tools to enhance career counseling and student success.
Previous studies have reported that auditory noise stimulation based on the stochastic resonance (SR) theory, which enhances the sensitivity of sensory systems, can improve auditory function. However, these studies used headphones for noise stimulation, and the convenience or practicality of the stimulation method has not been investigated. Therefore, this study aimed to investigate the effects of noise stimulation via bone conduction, which can be used without occluding the ear canal, on auditory function. Ten healthy adults participated in the study. Noise intensities were set at 40%, 60%, 80%, 100%, and 120% of each participant’s auditory threshold level. Auditory perception thresholds for weak sounds obtained from three-interval-three-alternative forced choice task were compared between the no noise (control) and the optimal noise stimulus intensity (OS) condition that maximizes the SR effect. The results showed that perception thresholds were significantly lower under the OS condition compared with the control condition. These findings suggest that noise stimulation via bone conduction, which offers high usability, can enhance auditory function.
End-user acceptance plays a key role in autonomous driving feature development. To maintain objectivity, autonomous functions should be assessed with key performance indicators (KPI) derived from physical parameters before they advance to the deployment stage. This study investigates the factors that prompt drivers to perform lane change maneuver and translates those findings into human readable explanations. Real-world highway data with manual lane change maneuvers was used. After preprocessing the data, an algorithmic pipeline automatically detected lane change events in the collected trajectory data and computed a wide set of physics based KPIs describing safety, comfort, and perceived safety. Model transparency was provided through SHapley Additive exPlanations (SHAP), ranking each KPI by its contribution to the lane change decision. The top-ranked SHAP features were then fed to a large language model (LLM) that generated concise natural-language rationales for every detected maneuver. By coupling accurate machine learning classification with explainable, driver-friendly feedback, the proposed framework aims to strengthen trust in autonomous lane change functionalities and promote their wider acceptance.
