Professor of Computer Science and UNESCO Chairholder, Secure High- performance Computing for Higher Education and Research, Namibia University of Science and Technology, Namibia
Friday April 10, 2026 10:00am - 10:10am GMT+07 Benchasiri 1Bangkok Marriott Hotel Sukhumvit, Thailand
Authors - Sunakshi Singh, Abhay Kumar Agrahari, Raghav Abstract - As cellular networks move toward 6G, traffic behavior becomes increasingly complex, shaped by user mobility and diverse service demands that vary across time and location. Accurate traffic prediction is therefore critical for efficient resource allocation and intelligent network operation. However, traditional statistical and conventional machine learning approaches rely on simplifying assumptions and struggle to capture the rich spatio-temporal interactions observed in large urban networks. Although recurrent models such as LSTM are effective at learning temporal patterns, they offer limited insight into how traffic evolves across geographically distributed regions. To address these limitations, this work frames cellular traffic prediction as a spatio-temporal learning problem and introduces a deep learning framework that jointly models temporal dynamics and spatial correlations using historical CDR data. The proposed approach is evaluated on real-world urban datasets and benchmarked against statistical and deep learning baselines, demonstrating superior prediction accuracy, faster convergence, and greater robustness under limited training data.
Authors - G.L.H.B. Gaweshika, T.G.I. Fernando Abstract - Optimization has become an active research area nowadays in every field majoring in Computer Science. This research focuses on developing an Optimal Personal Study Plan (PSP) generation system utilizing Metaheuristic Algorithms, considering the specific requirements of an individual student for a degree program. The PSP generation problem can be considered as an NP-hard problem, highlighting the need for efficient meta-heuristic algorithms to tackle this optimization challenge. The novel contribution of this work lies in the de-sign of a Genetic Algorithm (GA) and a Hybridized Genetic Algorithm-based Firefly Algorithm (GA-FA) for the PSP generation. The developed metaheuristic-based approach presents a promising avenue for enhancing the personalized study plan concept for students and academic support systems.
Authors - Kalpesh Popat, Divyakant Meva Abstract - Context Edge computing allows for processing data in real-time closer to its sources, which helps in applications like IoT, smart cities, healthcare, and industrial systems. However, security and privacy concerns hinder its mass adoption. This bibliometric analysis deals with security and privacy research in edge computing from 2023 – 2025. In compliance with the PRISMA guidelines, we con-ducted a bibliometric analysis on 643 peer-reviewed journal articles obtained from Scopus, employing methods such as analysis of publication trends, key-word co-occurrence, technology mapping, and domain analysis using VOSviewer and Biblioshiny software. Number of publications also grew exponentially (165 in 2023, 402 in 2024, 76 in early 2025 alone). The dataset provides h-index of 18 and g-index of 32. Security technologies such as blockchain, federated learning, and machine learning are prevalent. Primary domains include IoT networks, healthcare, and vehicular computing. All the publications are open access. Output in publications is led by China, India and the United States. This field shows fast maturing with the focus on lightweight cryptography, privacy-preserving mechanisms, and integration of the emerging technologies. Future research needs to focus on scalability, energy efficiency, and standardization to support mainstream adoption.
Authors - Divyakant Meva, Kalpesh Popat Abstract - Based on the total of 274 publications from 2023-2025, this bibliometric analysis reveals the evolving trend for agentic AI. This research utilized the PRISMA protocol and the Scopus database. From the results, there has been an exponential increase, which indicates that there has been a massive jump in the number of publications, specifically that there has been a “342% increase in 2025 from 2024.” Key results indicate that there has been intensive application in the fields of healthcare, education, and manufacturing, which comprise 18.2%, 14.6%, and 12.4% respectively. The United States has published the most, specifically at 38.7%, followed by China and European countries, which comprise 22.3% and 24.1% respectively. Thematic analysis Six major clusters emerged from the thematic analysis: autonomous systems, human-AI collaboration, ethical frameworks, multi-agent architectures, application in various domains, and evaluation methods. The study has demonstrated the shift from generative passive AI to autonomous agentic systems, identified important research gaps and presented future research directions.
Authors - Seungmin Lee, Ju-Won Park Abstract - The module-based static operating environment, which is widely used in domestic and international supercomputer operating centers, encounters numerous problems in supporting artificial intelligence / machine learning (AI/ML) parallel workloads because the variety of platforms and packages used make it difficult to build all execution environments. To address these issues and dynamically provide diverse execution environments, container-based cloud technologies are being widely utilized in high-performance computing (HPC) cluster systems. However, container runtime toolkits like Shifter and Singularity, which are widely used in the HPC field, present problems, such as the need for image format conversion, writing scheduler job script files, environmental setup, and direct management of the container lifecycle. This study proposes a solution to these problems by utilizing Kubernetes, which has become the de facto standard for container orchestration as it supports AI/ML parallel workloads even in HPC environments. Supporting Kubernetes-native parallel workload execution offers several advantages. First, image conversion is unnecessary because it directly uses Docker images. Second, human errors are minimized because the operator automatically handles the environment setup required for parallel execution. Third, in case of failures, automatic recovery and re-execution are possible by leveraging Kubernetes’ powerful container lifecycle management capabilities. In addition, this study introduces the distributed learning function of the KISTI Supercomputer web portal (MyKSC), which has been implemented using the proposed method.
Authors - Satish Kamble, Surendra Mahajan, Lalit Patil Abstract - In today’s world, IoT devices interact with each other for a specific purpose. IoT de-vices are used in every aspect of our lives. In IoT networks, devices can act as malicious nodes and can perform attacks affecting the IoT network's performance. A trust management system can play a major role in these IoT networks. This paper suggests a trust management system that is based on quality of services (QoS) and implemented on a real Raspberry Pi and ESP32 IoT testbed. The model uses direct trust and indirect trust parameters. The model uses memory efficiently by using sliding-window mechanism. This system implements a threshold-based mechanism for detecting untrustworthy devices and further blocking them for future communication. A recency-weight is used for stabilizing the system. This system is capable of detecting attacks such as the grayhole attack and RTT inflation attack.
Authors - Pitchayapatchaya Srikram, Thanapak Khattiya, Pathompong Charoansrimuang, Chayanit Yoosri, Nachirat Rachburee Abstract - Chalkiness in Thai Hom Mali rice is not only an important quality attribute for their market value and consumer acceptance, but also for rice grain breeding. However, conventional chalkiness evaluation relies on manual inspection, which is subjective and time-consuming. This study proposes an automatic multi-level chalkiness analysis framework based on semantic segmentation using a U-Net architecture with a ResNet34 encoder to segment rice grains and chalky regions from digital images. Then it estimates the grain counts for pixel-level segmented rice regions and chalky regions to classify chalkiness levels. We compare experimental results across datasets with and without the black background label. Both results are not significantly different in loss value, Mean IoU, Dice score, and F1 score. From a practical perspective, the segmentation of both datasets differs between rice and chalky regions due to illumination. The dataset, including the black background label, shows clearer chalky-grain segmentation regions and is closer to the ground truth. In contrast, the dataset excluding the background label shows chalky-grain segmentation regions and is closer to the original image.
Authors - Prashant Gaidhane Abstract - The control of robotic arms presents signicant engineering challenges due to their multi-input multi-output characteristics, strong coupling e ects, and inherent nonlinearities. The optimization landscape for controller parameter tuning exhibits multiple local optima, complicating the search for globally optimal solutions. Achieving precise end e ector path prole following in robotic systems demands sophisticated control methodologies tailored to handle these complexities. This re- search introduces an innovative cooperative foraging-based Grey Wolf Optimizer (CFGWO) algorithm to address these control challenges. The proposed methodology employs CFGWO to optimize the parameters of a PI D-based fuzzy regulator, targeting enhanced end e ector path prole performance in a Planar dual-link robotic arm with terminal load. The PI D-based fuzzy regulator incorporates additional design parameters beyond conventional PID structures, o ering expanded exibility in controller synthesis. The optimization performance of CFGWO is bench- marked against established algorithms including standard GWO, GWO- ABC hybrid, and LGWO variants. Performance evaluation focuses on minimizing the Integral of Time-weighted Absolute Error (ITAE) criterion. Results indicate that CFGWO achieves superior optimization con- vergence rates and delivers the lowest ITAE values among tested algorithms. Comprehensive experimental validation and performance analysis conrm the enhanced e ectiveness of the CFGWO approach, demonstrating its capability to balance exploration and exploitation mechanisms for robust global optimization in engineering applications.
Authors - Nandini Babbar, Anshika Shreshth, Saswati Gogoi, Sunil Kumar Abstract - Early and precise detection of skin cancer is very necessary, as it is one of the most aggressive diseases in the world, and its effective treatment is required. Because many skin cancer types appear visually similar and the available datasets are imbalanced, accurate diagnosis of skin lesions remains difficult using current medical technologies. Melanoma, one of the most severe skin cancer diseases, has a very low survival rate. In this paper, a multimodal is developed for classifying skin cancer by combining saliency maps with EfficientNetB3.This research work uses PAD-UFES-20 dataset to access and train the model. The clinicians can understand the lesion better through saliency maps, as they provide insightful information about the model’s decision-making process. This work concludes how deep learning models can be useful in improving skin cancer classification using an efficient approach for early detection clinically.
Authors - Vittorio Kuonadi Karimun Lie, Farrell Prema Tody, Gabriel Rinaldy Sudarmawan, Tiurida Lily Anita Abstract - The integration of biometric authentication technologies into smart hospitality environments introduces new challenges related to usability, privacy, and trust. This study evaluates biometric room access systems from a Human–Computer Interaction (HCI) perspective, focusing on how perceived security, perceived utility, perceived privacy, and perceived ease of use influence guest experience through trust. A study design that is quantitative was employed, and data were collected from 150 hotel guests who had previously used biometric room access. Partial Least Squares Structural Equation Modeling (PLS-SEM) was used to assess the suggested model. The results indicate that the model ex-plains 63.8% of the variance in guest experience and 59.1% of the variance in trust. Trust emerges as the strongest predictor of guest experience, while perceived privacy and perceived security significantly influence experience indirectly through trust mediation. In contrast, usability-related factors demonstrate comparatively smaller effects once baseline functionality is achieved. These findings suggest that biometric authentication in smart environments operates as a trust-sensitive socio-technical system, where perceived data governance and psychological assurance are critical determinants of experiential evaluation. The study contributes to intelligent systems research by demonstrating that authentication technologies embedded in physical access control contexts must integrate technical robustness with perceptual trust-building mechanisms to achieve sustainable user acceptance.
Authors - Sourabh Chordiya, Subhrakanta Panda, Akanksha Rathore Abstract - The rapid advancement of Artificial Intelligence (AI) and Large Language Models (LLMs) has unlocked powerful new capabilities for solving complex, multi-step problems. However, this progress has intensified concerns about the environmental sustainability of AI systems. While prior research has examined carbon emissions associated with training and inference in conventional LLM pipelines, emerging paradigms such as Agentic AI, where autonomous agents coordinate to execute multi-stage tasks, and Retrieval-Augmented Generation (RAG) introduce additional layers of computation that remain insufficiently studied from an emissions perspective. In particular, existing carbon measurement frameworks do not adequately capture the dynamic, distributed, and memory-intensive operations characteristic of these systems. This paper analyzes the limitations of current carbon accounting tools and available literature when applied to Agentic AI and RAG-based architectures. The widely used measurement frameworks capture only a fraction of the total computational footprint in such systems, largely omitting emissions arising from memory access patterns, retrieval processes, and inter-agent communication. These overlooked components become increasingly significant as AI workflows shift from single-system inference toward multi-agent orchestration and knowledge retrieval pipelines. Based on this analysis, the paper proposes directions for a comprehensive life-cycle carbon assessment framework and an Eco Rating tailored to next-generation AI systems. Such a framework must account for heterogeneous hardware usage, dynamic inference paths, retrieval infrastructure, and communication overhead across distributed agents. The findings highlight a substantial blind spot in current sustainability evaluations and underscore the urgent need for standardized methodologies that reflect the true environmental impact of emerging AI paradigms.
Authors - Kaja Mohideen A, Senthil Prakash PN Abstract - Breast tumor segmentation using mammographic is a difficult task because mammographic images have low contrasts, complex tissue structures, and high inter patient variability. Radiologists commonly make left-right-breast comparisons to detect suspicious inconsistencies in the image of the left and right breast in the routine clinical practice. It is based on this bilateral diagnostic strategy that this paper suggests a difference-guided bilateral U-Net to inter pretable breast tumor segmentation. Paired left and right mammogram of the same patient are first adjusted by the horizontal flipping and intensity normali zation. A pixel-based difference image is then created to highlight disparities that are absolutely in nature to highlight areas that are asymmetric and which might reflect pathological alterations. To make the network learn both appear ance-based and asymmetry-driven representations, the bilateral mammograms are proposed to be jointly processed with the respective difference map, after which the network will be trained. This design enhances the performance of segmentation without compromising clinical interpretability because it explicit ly points out areas of interest. The suggested method is tested on publicly ac cessible data, such as MIAS and CBIS-DDSM and real-time mammographic images obtained in a clinical setting. The experimental data indicate that differ ence-guided framework provides higher segmentation accuracy and lower false positive rates than single-breast U-Net models, which implies that the frame work can be used to delineate breast tumors on automated mammography.
Authors - Subham Ghosh, Banani Basu, Arnab Nandi Abstract - Radio-frequency based human activity recognition (HAR) using wearable antennas has recently gained interest due to its promise for comfortable and effective monitoring in applications such as smart healthcare and surveillance. However, traditional deep learning (DL) models for HAR are often constrained due to their reliance on large datasets and poor generalization performance. This paper presents an innovative framework for capturing and recognizing two-hand movements by using the near-field of a wearable antenna. The proposed system innovatively integrates signal smoothing, Morlet wavelet transform (MWT) time-frequency (TF) transformation, feature extraction based on statistical significance using the Kruskal-Wallis test, and a quantum artificial neural network (QANN) for robust feature learning and classification. The performance of the suggested technique is systematically compared against traditional machine learning models. Experimental results demonstrate that the proposed framework achieves superior classification performance for hand activity identification, underscoring its efficacy and promise for wearable RF-based HAR systems.
Authors - Mohd Hizami Ab Halim, Suriani Mohd Sam, Norliza Mohamed, Hazilah Mad Kaidi, Norulhusna Ahmad Abstract - Accurate identification of fish species based on morphometric and meristic characteristics is challenging, particularly for commercially important species such as Megalaspis Cordyla, due to subtle morphological differences and limited labelled data. This review examines recent advances in deep metric learning, with a focus on Siamese network architectures, for few-shot morphometric and meristic identification of M. Cordyla. We synthesize studies on metric-based similarity learning, landmark-driven morphometric analysis, and finegrained fish classification to show how Siamese networks effectively learn discriminative embedding spaces under low-data conditions. The review also analyzes reported performance comparisons across the literature, including classification accuracy, precision-recall behavior, robustness to small training sets, and generalization to unseen species or populations. Overall, the findings indicate that Siamese and deep metric learning-based approaches consistently outperform conventional classification models in fine-grained fish identification tasks, while highlighting open challenges such as the lack of standardized morphometric datasets for Megalaspis Cordyla, limited meristic-aware benchmarking, and the need for interpretable similarity measures to support fisheries science and biodiversity conservation.
Authors - Aniket Chatterjee, Anirban Dasgupta, Parvez Aziz Boruah, Raktim Acharjee Abstract - Underwater gesture detection is a well-known area of research in recent times that helps in communication between divers and Underwater Remotely Operated Vehicle (ROV). Hand gestures are commonly used in underwater environments as a straightforward and intuitive method for conveying commands or messages between divers and ROV. The ROV need to first detect and identify the human and then detect his/her hand and what type of gesture it is. However, the underwater environment has many challenges: turbulent waters can disrupt the ROV navigation and obstruct the capture of clear video footage, resulting in noisy images that complicates the accurate recognition of hand gestures. Besides that, the ROV must process visual data and respond quickly, especially in critical situations where quick decision making is required. This project work aims to optimize the ROV application program for improved real-time image processing and gesture recognition, that helps in effective communication even under challenging underwater conditions. Six different models have been explored including techniques like Channel Attention Mechanism and Spatial Attention. Our developed model(UnderwaterGestureNet) have shown better result with less number of parameters. This lightweight model is more efficient to deploy in embedded system of an ROV.
