Authors - Nikhil Kumar, Anurag Barthwal, Shakti Kundu Abstract - Falls from an altitude are among the most common causes of both fatal and non-fatal injuries in the global community and second only to road traffic accidents in accidental mortality. One of the primary problems in alleviating the effects of such incidents is the late detection and reporting of falls, especially in the cases where witnesses are not present, which exposes the victim to a high risk of severe injuries, or even death, because of the lack of medical care. To curb this problem, this paper proposes an effective and affordable smartphone-based solution towards automated detection of human falls off heights. The suggested solution uses built-in smartphone sensors namely accelerators and barometers to record motion dynamics and changes in altitude which are linked to falls. The primary characteristics, such as the absolute linear acceleration, change in altitude, are acquired and applied to train and test a Support Vector Machine (SVM)-based classification model, which shows strong performance, with the F1-score of 0.94, which, in turn, proves the high reliability of the model in differentiating between fall and non-fall events. The results indicate the success of the multi-sensor data fusion with machine learning methods and emphasize the possible relevance of the given system to practical applications in the field of fall detection in real-time, early emergency response, and the overall occupational and population safety schemes.
Authors - Van-Cuong Nguyen, Huu-Cuong Nguyen, Quang-Hieu Ngo, Trong-Hieu Luu, Thanh-Tam Nguyen Abstract - This paper aims to introduce the relationship between surface roughness and rice yield on paddy field using camera mounted on UAV. Unlike other studies where people focus on genes and rice varieties, we think that the surface roughness also has a big impact on rice yield. We surveyed paddy by using successive aerial images, generated the ortho-photos before conducted the surface roughness by using Harralick texture extracting. From the resulting mapping photo, we chose three distinct local areas for sample data collection based on the surface differences. Three different treatments were applied across these areas, with agronomic traits and yield components meticulously documented. As the crop season progressed, discernible disparities in crop vitality emerged, observable both in the field and through analysis using the Normalized Difference Vegetation Index (NDVI). Furthermore, our rigorous evaluation of agronomic traits and yield components revealed statistically significant disparities among treatments, reaching the remarkable 1% significance level. These findings hold considerable promise for farmers, facilitating informed decisions in land use planning for subsequent crop seasons.
Authors - Abhay Kumar Agrahari, Snehal Rajput, Omji, Akhil Pandey, Chiluka Varshith Reddy Abstract - In today’s era, reliable and safe communication has become a major requirement in smart vehicle networks. In this research work, we present a specific method for authentication between the vehicle’s on-board unit (OBU) and roadside unit (RSU), which uses Physical Unclonable Function (PUF). This technology provides an identity for each vehicle unit that cannot be repeated. In this process, both units are registered with a reliable authority, which is the basis of certification. The process of mutual certification not only pays attention to safety, but has also been made faster with minimal resources. The validation of the protocol is checked via the ROR model and the AVISPA tool, which shows that this model is protected from common security threats. In addition, we will compare our proposed protocol with predefined algorithms on the basis of communication cost and also do the security analysis. This study offers a general description of the VANET authentication system that is practical, safe and skilled.
Authors - Mohd Mansoor Khan Abstract - An exclusive action dataset, termed the ImuFall, was created using gyroscope data from the MPU6050 IMU sensor. An end-to-end posture and fall detection system was developed and evaluated on this dataset. A threshold-based mean slope algorithm was implemented and compared with machine learning methods, namely ν-SVM for posture classification and random forest classifier (RFC) for fall detection. The ν-SVM was chosen to reduce overfitting, while RFC was used for its effectiveness with time-series data. The cascaded framework achieves 100% best-case accuracy, with 95.8% average posture accuracy and 100% fall detection accuracy. This is the first reported implementation of a cascaded ν-SVM–RFC end-to-end fall detection system.
Authors - Maya Fitria, Muhammad Hafiz Rinaldi, Khairun Saddami, Isack Farady, Kahlil Muchtar, Sayed Muchallil Abstract - As the most consumed commodity worldwide, banana requires careful and proper growth management to maintain its production, including maintaining its leaf health. Commonly, farmers identify the disease in banana leaves by inspecting its appearance. However, this conventional method is considered subjective to one person to another, and this could lead to delayed treatment, and may impact the fruit development and production. To address this issue, this re-search proposed B-Leaf Scanner, a mobile-based application integrating a deep learning approach for banana leaf disease detection. The application integrated the YOLOv5-based model to detect and classify the disease in banana leaf which is conducted by capturing image from a camera or by inputting from the device gallery. The proposed application was designed aligned with the findings from field observations and interviews with local farmers to ensure usability and related to real-world settings. The findings show that the detection model yielded an mAP of 80.1%, following with 86.8% and 72.4% of precision and recall value, respectively. These results indicate the reliability of the model in performing the detection process. Moreover, the usability testing of the application was con-ducted to ten local farmers through task-based testing, and System Usability Scale (SUS). Based on usability results, the B-Leaf Scanner application achieved excellent usability with a SUS score of 88%, indicating the application can effectively support local banana farmers in identifying leaf diseases.
Authors - Sanjeeb Prasad Panday, Ujawal Thapa, Basanta Joshi, Aman Shakya, Anunaya Pandey Abstract - Early diagnosis of colorectal diseases depends upon the detection of polyps in colonoscopy images. These polyps often blend into their surrounding which often poses a challenge in detecting them. In this regard, we introduce a new approach that improves polyp segmentation using distraction mining. Our method is based on the enhancement of Positioning and Focus Network (PFNet) which was originally designed for camouflaged object segmentation. The PFNet first identifies potential polyp regions using the Positioning Module (PM) and then refines the detection by focusing on hard-to-distinguish areas using the Focus Module (FM). We integrate a distraction mining technique into FM which helps the model differentiate polyps from misleading background details and further improved the accuracy. The comparison of the PFNet model with other models like SINet and PRANet. The PFNet models and other models like SINet and PRANet are evaluated on a different polyp datasets like Colon DB, Laribpolyp DB, and CVC-300. The result shows that the distraction mining enhance the segmentation performance on a complex datasets like laribpolyp DB with 0.8046 for S-measure, 0.6651 for weighted F-measure,0.0202 for MAE,0.8590 for adaptive E-measure, and CVC-300 with 0.8220 for S-measure, 0.7317 for weighted F-measure, 0.0299 for MAE and 0.8735 for adaptive Emeasures. There are slightly low accuracy in the colon DB datasets.
Authors - Jose Alejandro Ascencio-Laguna, Armida Gonzalez-Lorence, Ana Lilia Mondragon-Solis, Victor Alberto Gomez-Perez Abstract - Machine Learning (ML) and geospatial clustering have traditionally been applied as independent approaches to urban freight transportation chal lenges, particularly arrival time prediction under "just-in-time" constraints. De spite their complementary nature, their integration remains underexplored, while distance-based methods relying on Euclidean metrics yield error margins of 18 35 minutes, insufficient for operational logistics. This study proposes a hierarchical framework combining geographic k-means clustering (k=14) as a spatial segmentation layer with an enhanced Random For est regressor incorporating temporal feature engineering. The architecture is com putationally efficient and robust to real-world uncertainty after training. The framework was validated across three metropolitan areas in Mexico using 306,847 records from June 2024, benchmarked against five algorithms through stratified temporal validation and Wilcoxon tests with Bonferroni correction. The proposed model achieved a Mean Absolute Error of 347.2 seconds (5.79 min), representing a 68.1% reduction relative to historical baselines (MAE: 1,089 s) and a 19.9% improvement over standalone Random Forest (MAE: 433 s). Eu clidean distance was the dominant predictor (43.7%), followed by geographic coordinates (32.8%). All improvements were statistically significant (p
Authors - Paresh Baidya, Rourab Paul, Vikas Srivastava, Sumit Kumar Debnath Abstract - Kyber is a lattice-based key encapsulation mechanism se lected for standardization by the NIST Post-Quantum Cryptography (PQC) project. A critical component of Kyber’s key generation process is the sampling of matrix elements from a uniform distribution over the ring Rq. This step is computationally intensive and significantly impact ing task in the performance of low-power embedded systems such as Internet of Things (IoT), wireless sensor networks (WSNs), smart cards, etc. Existing approaches like SampleNTT and Parse-SPDM3 rely on rejec tion sampling, need at least three SHAKE-128 squeezing steps per poly nomial. As a result, it causes significant amount of latency and energy. In this work, we propose a novel and efficient sampling algorithm, namely Modified SampleNTT, which substantially reduces the average number of bits required from SHAKE-128 to generate elements in Rq—achieving approximately a 33% reduction compared to conventional SampleNTT. Modified SampleNTT achieves 99.16% success in generating a complete polynomial using only two SHAKE-128 squeezes. Furthermore, our algo rithm maintains the same average rejection rate as existing techniques and passes all standard statistical tests for randomness quality. FPGA implementation on Artix-7 demonstrates a 33.14% reduction in energy, 33.32% lower latency, and 0.28% fewer slices compared to SampleNTT.
Authors - Kingston Pal Thamburaj, Ramesh Mercedes Premalatha, Mukhlis Abu Bakar Abstract - Large language models are increasingly used to generate educational explanations, but hallucinations, uneven language quality, and untraceable confidence can introduce misconceptions. These risks are amplified in bilingual classrooms, where meaning must remain aligned across languages and low-resource language support is limited. This paper introduces a trust-aware multi-agent validation architecture for bilingual Tamil-Malay AI-generated educational content. The architecture decomposes validation into specialized agents that verify factual claims via evidence-grounded retrieval, assess linguistic well-formedness and terminological consistency, estimate pedagogical suitability for a target grade level, detect hallucination and bias risk, and measure cross-lingual semantic consistency to identify drift between Tamil and Malay explanations. Agent outputs are combined through a transparent aggregation mechanism to produce an overall bilingual trust score and an interpretable validation report with actionable revision cues. A benchmark construction protocol and evaluation methodology are presented to quantify claim-level correctness, cross-lingual agreement, and trust-score calibration against expert annotations. The proposed approach supports human-AI collaborative content authoring and intelligent tutoring workflows, improving the reliability and inclusiveness of bilingual education systems in Southeast Asian contexts.
Authors - Komendra Sahu, Mallikharjuna Rao K., Sonali Agarwal Abstract - This study examines whether textual complexity in corporate disclosures predicts stock excess returns. Building on prior research using Loughran–McDonald (LM) tone variables, the baseline ordinary least squares (OLS) results are replicated and the analysis is extended in three directions. a novel Corporate Communication Text Complexity Index (CCTI) is developed using structural and linguistic features of SEC 10-K and 10-Q filings. market-based controls, including volatility and momentum, are incorporated. machine learning models are applied to capture potential nonlinear dependencies. Analysis of a large sample of filings from 2009 to 2024 demonstrates that OLS models have near-zero explanatory power, consistent with previous findings. In contrast, Random Forest models significantly improve predictive performance (R2 = 0.19944), indicating that excess returns are influenced by nonlinear patterns in textual complexity. Polynomial regression also reveals a convex relationship, with extreme textual complexity associated with negative excess returns. Analysis of a large sample of filings from 2009 to 2024 confirms that OLS exhibits near-zero explanatory power. This finding is consistent with prior research. In contrast, Random Forest models substantially improve predictive performance (R2 = 0.19944), indicating that excess returns respond to nonlinear patterns in textual complexity. Polynomial regression reveals a convex relationship, where extreme textual complexity is associated with negative excess returns. Overall, these results indicate that market reactions to complexity are inherently nonlinear and cannot be adequately captured by traditional tone-based linear models.
Authors - Gunchita Kaur Wadhwa, Rugved Dinesh Kshirsagar Abstract - Increasing infrastructure structures are being exposed to outdoor environmental factors such as UV, water, humidity, temperature fluctuations and air pollutants. At the same time, increasing trend of smart cities is highly dependent on successful implementation of wireless sensor networks to be able to measure e.g. the intensity of UV, air quality, temperature and humidity. Therefore, this research focusses on developing multifunctional nanocomposite coating composed of BNNS dispersed in PVA deposited on PMMA transparent panels that provides an efficient solution to many challenges related to smart structure infrastructure. This research demonstrates a coating material that, after optimizing its structural properties, behaves as following in one step solution: (i) effective UV shield using boron nitride nanosheets as filler, (ii) exhibiting superhydrophobic self-cleaning properties for water and chemicals after structure modification and chemical surface treatment, (iii) acting as an atmospheric energy harvester by using the tribocatric effects between the coating and raindrops for charge extraction, and (iv) behaving as micro-scale energy storage due to dielectric characteristics of BNNS within the coating, which could be potential to power Internet of Things (IoT) low power consumption sensor nodes. The multifunctional coating therefore represents a new class of self-powered smart-city surfaces capable of protecting infrastructure materials while simultaneously harvesting and storing environmental energy. The proposed approach contributes to sustainable urban development and aligns with Sustainable Development Goals related to clean energy and resilient cities.
Authors - Arathi Kumaresan Chandirakala, Sunantha Sodsee Abstract - Synthetic kidney image augmentation plays critical role in improvising quantity and diversity of health imaging data. But anatomic generation of visually realistic synthetic images remains as a major challenge, often resulting in poorer texture quality, mode collapse, and loss of structural details. Existing approaches frequently struggle to preserve consistency in texture, shape, and intensity alterations, limiting their effectiveness in clinical applications. To tackle these limitations, the Adaptive Schrodinger Optimizer enabled Deep Convolutional Generative Adversarial Network (ASRA_DC-GAN) is proposed for augmenting synthetic kidney image. Initially, input kidney Computed Tomography (CT) image is categorized as majority and minority class. Further, image enhancing separation among elements is performed for both classes by Histogram Equalization. Further, augmentation of synthetic kidney image is done through DC-GAN in case of minority classes. Herein, DC-GAN is tuned by ASRA, which is formed by combination of Adaptive concept and Schrodinger Optimizer (SRA). Finally, the attained outputs are allowed for generation of augmented new balanced dataset. Performance of proposed ASRA_DC-GAN is assessed by Second-Derivative like entropy and Measure of Enhancement (SDME), which gained outstanding values of 0.839 and 46.90dB.
Authors - Komendra Sahu, Aayush Sahu, Aparajita Vaish, Kavita Jaiswal Abstract - The AWARE framework (USENIX ATC ’23) applied meta learning so reinforcement learning (RL) agents could adapt more quickly to different workload patterns. However, this approach still assumes that workloads seen during deployment are similar to those used during train ing. When this assumption breaks, system performance can decline. In the real world, workload behavior often changes due to traffic spikes, configuration updates, or shifts in resource demand. Under these condi tions, a fixed meta-policy may no longer reflect the current environment, leading to unstable scaling decisions. To handle this , we introduce a Shift-Aware Meta-PPO framework. The system tracks workload behav ior using the KL-divergence to detect changes in distribution. When a shift is detected, the meta-buffers are cleared and exploration resumes, allowing the RL agent to adjust its policy to the upcoming new work load. Tests show that this approach stays stable during workload changes and avoids the sharp performance drops seen in standard meta-learning methods under out-of-distribution (OOD) workloads.
Authors - Atrey Kantharaj Urs, Madhan Kumar Srinivasan Abstract - The proposed work presents a Mixed Reality (MR) shopping system designed to address persistent challenges in online fashion retail, including fit uncertainty, limited personalization, and the lack of immersive experiences, by integrating real-time virtual try-on, avatar-based visualization, and an AI-powered recom mendation engine. The system allows users to explore and evaluate garments as interactive three-dimensional models within their physical environment, thereby improving confidence in style and fit decisions. A hybrid recommendation frame work combines body-feature matching, content-based and collaborative filtering, contextual interaction signals, and foundational fashion design principles to gen erate personalized outfit suggestions, while an AI assistant delivers explainable recommendations and interactive guidance throughout the shopping journey. By effectively bridging the gap between physical retail and digital platforms through adaptive AI models and MR visualization, the system offers a practical alter native to conventional online shopping, demonstrating the potential of Mixed Reality to create a more immersive, intelligent, and user-centric fashion shopping experience that enhances decision-making, increases engagement, and reduces product returns.
