DAY 2 (29 Oct 2025)
Track 5 (11:30am-1:00pm)

Computing and Computational Intelligence (CCI) 1

Track B5F1 CCI 5.1: Computing & Computational Intelligence (CCI) 5.1

Room: F1. Sipadan I (Level 4)

Chair: Jamal Ahmad Dargham (Universiti Malaysia Sabah, Malaysia)

11:30 Securing Face ID: Privacy Preservation for Non-Retentive Face Recognition System

Megan Chua, Chanelle Yue-Ting Yeow, Cassandra Xin-Yee Chwee and Shu-Min Leong (Monash University, Malaysia Campus, Malaysia); Raphael C.-W. Phan (Monash University, Malaysia)

Facial recognition technology is increasingly integrated into various applications. While face recognition systems have streamlined the authentication process and reduced the need for manual verification, the advancement in artificial intelligence (AI) generating realistic-looking media raises significant privacy concerns due to the potential misuse of biometric data. While current security protocols ensure data protection, storing biometric data in the system is a latent risk. In cases where the stored data is compromised, the users are susceptible to attacks such as face swapping, deepfake and identity theft. To address this, this paper presents a privacy-preserving algorithm that omits the need to store human raw biometric data in the system. This is achieved by utilizing a new combination of Locality-Sensitive Hashing (LSH), salting, and RSA encryption for face recognition. The proposed method ensures data security by securely hashing and encrypting facial features while maintaining high recognition accuracy. The proposed framework is evaluated on the Labeled Faces in the Wild (LFW) and achieves a comparable performance with the state-of-the-art techniques.

11:45 The AI Data Analyst: A Framework for Autonomous Data Analytics, Highlighting LLM and AI Agents

Vichayada Laosubinprasert, Proadpran Punyabukkana and Atiwong Suchato (Chulalongkorn University, Thailand)

This work presents an automated data analytics framework that integrates large language model (LLM) and artificial intelligence (AI) agents to perform end-to-end analysis based on Google's six-step methodology: Ask, Prepare, Process, Analyze, Share, and Act without human involvement throughout the analytics process. The system requires only user input containing the objective, data, context, and prior hypotheses. Then, the system autonomously generates prompts and executes tasks through specialized AI agents. AI agents perform the roles of planning tasks and directing the actions of agents, leveraging LLM for generation of ideas and reasoning to inform their plans and actions. Experiments on two datasets across domains, including education, and business show strong performance. Average analytics scores range from 8.6 to 9.4 out of 10, with average execution times varying between 1.9 and 6.6 min, and error rate varying between 0.2 and 1.2 occurrences per run. Additionally, the system can perform complex tasks such as coding, statistical analysis, machine learning modeling, and visualization generation. The results demonstrate the potential of LLM to function as virtual data analysts, enabling fully automated, domain-independent analytics.

12:00 An Ensemble Clustering Approach to Recognize Flight State in Pilot Training

Anam Iqbal and Graham Wild (University of New South Wales, Canberra, Australia)

The aviation industry has been growing rapidly, and the success of the sector depends heavily on pilots receiving rigorous training, which guarantees the safety of air travel. During various flight states of an aircraft, pilots encounter different situations; to anticipate and manage potential hazards, pilots must possess detailed knowledge of these flight regimes and maintain appropriate control. This study employs an ensemble clustering approach on flight data simulated through Monte Carlo Simulations to determine and recognize crucial flight states of an aircraft, providing the basis for assessing pilot performance within each state. The clustering performance of various individual algorithms and their combinations is evaluated. Among single clustering models, Shared Nearest Neighbor (SNN) achieved the highest performance measures, while ensemble clustering models further enhanced clustering performance, with the combination of SNN and Fuzzy C-means performing particularly well. When comparing individual and ensemble approaches, it is observed that ensemble methods detect and distinguish flight states more accurately than any single clustering algorithm. This research marks a significant advancement in the use of ensemble approaches to improve the diagnosis of pilot performance during specific flight states, thereby enhancing training and flying safety.

12:15 Attack-SH: Adversarial Attacks on Self-Healing Material Properties Prediction Model

Min Xuan Tan (Monash University Malaysia Campus, Malaysia); Pei Sze Tan (Monash University, Malaysia Campus, Malaysia); Raphael C.-W. Phan (Monash University, Malaysia); Shu-Min Leong (Monash University, Malaysia Campus, Malaysia)

AI is now increasingly applied in diverse domains. The recent Nobel prizes for Physics and Chemistry awarded to computational scientists shows the significant impact that AI has on real-world scientific applications. Adversarial attacks pose a significant threat to the reliability of AI systems, particularly in high-stakes applications such as those in the materials sciences domain, which affect interactions with materials that exist in the real world. This paper examines the impact of two widely used adversarial attack approaches, notably the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), on a target AI model's performance for real-world materials. Experimental results demonstrate that both methods effectively degrade predictive accuracy, with PGD showing a more severe effect. Notably, high Structural Similarity Index (SSIM) scores across all perturbed samples suggest that the attacks introduce imperceptible changes, increasing their potential risk as such attacks are then undetectable. Further analysis using metrics such as Mean Squared Error (MSE), Adversarial MSE (A-MSE), and Relative Error Increase (REI) confirms substantial shifts in model output and reconstruction quality. These findings highlight critical vulnerabilities in current model architectures and emphasize the urgent need for more resilient defense strategies, such as adversarial training and input pre-processing.

12:30 Deepfake Detection Using ResNet50V2 with Machine Unlearning Integration

Md Serajun Nabi (Multimedia University, Malaysia); Dema Yuden (Albykhary International University, Malaysia); Mohammad Faizal Ahmad Fauzi (Multimedia University, Malaysia)

The advent of deepfake content has been an enormous challenge in digital media security that necessitates an effective detection system. This study proposes a deepfake detection model based on ResNet50V2 with an associated machine unlearning (MUL) approach to enable selective forgetting of generated data. The model is trained and evaluated on the FaceForensics-1600 dataset, which consists of real and deepfake video frames, and then undergoes a retraining phase, excluding the forgetting set. Performance before and after unlearning is quantified in terms of classification reports, confusion matrices, and ROC curves. Experimental results show that the model maintains a steady accuracy of 96% and ROC-AUC of 0.97 even after unlearning. The findings suggest that MUL can complement model adaptability in dynamic data environments. It also supports compliance with data privacy laws requiring data deletion. The results demonstrate that practical unlearning can be applied to deepfake detection systems without impacting performance, offering a promising solution for ethically adaptive and privacy-preserving AI.

12:45 Preserving Cluster Identity Across Time: an Incremental Cosine Similarity Approach

Adarsh Suresan Nair (Yum India Global Services Private Limited, India); Aishwariya K K (IBM India Private Limited, India)

In many real world applications, such as customer behavior analysis, data distributions evolve gradually over time.Traditional clustering methods, which retrain models from scratch at regular intervals, often fail to preserve the fine grained dynamics within clusters as they evolve. In this paper, we propose an incremental clustering framework that maintains cluster continuity across sequential time periods. Our approach clusters new incoming data independently and then maps the resulting clusters back to existing clusters by computing cosine similarity between cluster centroids. If a sufficiently similar match is found, the cluster is continued; otherwise, a new cluster is initialized.This method enables accurate tracking of cluster evolution overtime, capturing subtle shifts in cluster characteristics without abrupt reassignments. Experimental results on synthetic and real-world datasets demonstrate the effectiveness of our approach in preserving cluster lineage and detecting emerging behaviors.Our framework offers a simple yet powerful solution for time-sensitive applications where maintaining historical cluster context is critical.

Control Systems & Robotics (CSR)

Track B5F2 CCI 5.2: Computing & Computational Intelligence (CCI) 5.2

Room: F2. 501 Kadamaian (Level 5)

Chair: Nordin Ramli (MIMOS Berhad, Malaysia)

11:30 Gesture-Driven Cursor Control Interactive Projection System

Lance Eman M Hernandez, Jhon Lester Malabuyo, Carl Symon V Ofrin, Rocxel Roi C Puso and Reniel M Cornejo (FAITH Colleges, Philippines)

The researchers introduced an interactive projection platform that enables users to control on-screen content through natural hand gestures, enhancing the way people interact with digital media. This innovative system integrates gesture recognition technology, high-definition cameras, and projectors to detect user movements in real time, allowing seamless navigation of presentations and educational materials. By eliminating the need for traditional remote controls and whiteboards, it fosters a more engaging and dynamic experience in both academic and professional environments. The system's core advantage lies in its precise gesture recognition and smooth performance, powered by advanced machine learning algorithms based on Google MediaPipe and supported by high-quality hardware components. Its modular design allows for quick setup, instant response times, and handheld functionality, ensuring adaptability across various settings. Cost-effectiveness was a key consideration, with extensive testing and refinement addressing detection accuracy, system compatibility, and environmental challenges. Ultimately, this gesture-based interface aims to transform presentation methods by promoting immersive, intuitive, and highly interactive user experiences.

11:45 Blockchain Networks Metrics Collection for Validation in Blockchain Simulator

Tze Yee Choo (Heriot-Watt University Malaysia, Malaysia); Timothy Tzen Vun Yap (Heriot-Watt University, Malaysia); Ian K.T. Tan (Heriot-Watt University, Malaysia & Innov8tif Solutions Sdn Bhd, Malaysia); Zi Hau Chin (Heriot-Watt University Malaysia, Malaysia)

The accuracy and reliability of the blockchain networks simulation depend on the realistic network and simulator configuration. However, many existing blockchain simulators may use outdated parameters that fail to reflect the dynamic and constantly evolving nature of real-world blockchain systems. Hence, a simulator validation is crucial before starting the process of improving blockchain performance. In this paper, we collect the recent blockchain-related data from different sources and re-parametrize the blockchain parameters. A different network setting, namely a 7 regions simulation, was proposed to be used in the blockchain simulator to improve the reliability of the blockchain simulator. We then compare the block propagation time between the proposed settings with the original settings and the results of the DSN network, which reflects the Bitcoin network using the rate of change rule. The results show that the proposed settings are slightly closer to the Bitcoin network compared to the original settings in terms of block propagation time. The results are then validated using the standard deviation to verify consistency. We also suggest further collecting the data for a longer duration to achieve higher reliability on the simulator.

12:00 Transformer and CNN-Based Lightweight Ensemble for Retinal Disease Classification Using OCT Imaging

Aditya Bhongade (Yeshwantrao Chavan College of Engineering, India); Yogita Dubey (Yeshwantrao Chavan College of Engineering, Nagpur, India); Punit Fulzele (Directorate of Research & Innovation, SPDC Datta Meghe Institute of Higher Education & Research Wardha, India)

Automated retinal disease classification plays a crucial role in the early detection, monitoring, and treatment planning of ocular disorders, ultimately helping to prevent vision loss. This paper introduces a robust and effective ensemble model that combines the strengths of ResNet-50 and Vision Transformer (ViT-B16) to classify retinal images from the widely used OCT 2017 dataset. The ensemble leverages the local pattern recognition and fine-grained feature extraction capabilities of ResNet-50, along with the global contextual understanding and long-range dependency modeling of ViT-B16. By integrating both convolutional and transformer-based paradigms, the proposed method enhances classification accuracy and generalizability. The model achieves state-of-the-art (SOTA) performance with an accuracy of 0.9959, an AUC of 1.00, and F1-score, precision, and recall of 0.9959, significantly outperforming several existing approaches. These results demonstrate the effectiveness and reliability of the ensemble for real-world clinical deployment, offering a promising solution for automated retinal disease diagnosis and advancing modern medical image analysis.

12:15 Hybrid FEM-ANN Modeling of Rutting in Flexible Pavements Reinforced with Coir Geotextile at the Subbase-Subgrade Interface

Sheina Pallega and Dante L Silva (Mapúa University, Philippines); Jimmy G. Catanes (Commission on Higher Education, Philippines); Mark B. Ondac (Mapúa University, Philippines)

The increasing demand for resilient and sustainable road infrastructure has led to the integration of innovative materials and advanced computational techniques in pavement engineering. This study investigates the rutting performance of flexible pavements reinforced with coir geotextile, a biodegradable material derived from coconut husk, offering an eco-friendly alternative to synthetic reinforcements. A finite element model (FEM) was developed in ABAQUS to simulate rutting behavior under static and dynamic loading, incorporating varying material properties and layer thicknesses. Model validation using experimental data yielded a MAE of 0.184, confirming its predictive accuracy. A total of 200 datasets were extracted from the validated FEM and used to train an Artificial Neural Network (ANN) for rutting depth prediction. The ANN, structured as a feedforward backpropagation model with 20 input parameters, one hidden layer with 41 neurons, and one output, achieved high correlation coefficients across all phases: training (R = 0.99051), validation (R = 0.96830), and testing (R = 0.99036), with an overall R of 0.98874 and MSE of 0.19331. Sensitivity analysis using Garson's algorithm ranked input parameters based on their relative influence, identifying asphalt elasticity and thickness, and subgrade stiffness as the most critical factors affecting rutting performance. This research offers a robust, data-driven approach to pavement design by integrating FEM, ANN, and sensitivity analysis, providing accurate rutting depth predictions and valuable insights for material optimization. The findings support the development of cost-effective and environmentally sustainable pavements, aligned with global sustainability goals.

12:30 Real-Time Vision Inspection with AI-Based Edge Processor for Printed Circuit Board Assembly Quality Control

Jun Le Lai (Swinburne University of Technology Sarawak, Malaysia); Hudyjaya Siswoyo Jo (Swinburne University of Technology Sarawak Campus, Malaysia)

In modern electronic manufacturing, particularly in printed circuit board (PCB) assembly, ensuring product quality through automated inspection is essential. While traditional automated optical inspection (AOI) systems are effective, they often face limitations when dealing with complex assemblies and are typically designed for specific board types, limiting flexibility and scalability. This research presents the development of a universal, real-time inspection system for PCB assembly using artificial intelligence (AI) integrated with edge computing. The system is trained on a diverse dataset containing common PCB assembly defects, such as missing components, solder bridging, and insufficient solder. A lightweight AI model is developed to enable efficient inference on low-power edge processors. The proposed system is evaluated in real-world testing scenarios using different PCB assemblies to validate its adaptability and feasibility for deployment in practical manufacturing environments. Results demonstrate the potential of AI-based edge solutions in enhancing the flexibility and performance of automated quality control processes.

12:45 Classification of Natural Disaster Images Using Convolutional Neural Network Models

Mahmoud Yehia Emam Selim Rehan and Noramiza Hashim (Multimedia University, Malaysia); Khairil Anuar (Multimedia Universiti, Malaysia); Wan Noorshahida Mohd-Isa (Multimedia University, Malaysia)

This paper explores the use of Convolutional Neural Networks (CNNs) for disaster classification, focusing on the EfficientNet architecture to classify four major natural disasters: floods, earthquakes, cyclones, and wildfires. EfficientNet stands out due to its novel compound scaling approach, which significantly enhances feature extraction from diverse image data while maintaining high computational efficiency. The model was trained using transfer learning on a carefully balanced dataset sourced from multiple disaster imagery repositories. Its performance was evaluated through accuracy, precision, recall, F1-score, and confusion matrices, ensuring a rigorous and reliable assessment of classification effectiveness. Experimental results demonstrate that EfficientNet consistently outperforms six competing models-VGG16, ResNet50, MobileNet, ARCNet-MobileNet, RescueNet, and ARCNet-VGG16. Notably, EfficientNet achieved the highest accuracy of 94% while requiring the shortest training time of only 13.5 minutes for 10 epochs. These findings highlight EfficientNet's scalability, robustness, and reliability, making it an excellent candidate for deployment in real-world, image-based disaster management and early warning systems.

Power, Energy & Electrical Systems (PES)

Track B5F3 PES 5: Power, Energy & Electrical Systems (PES) 5

Room: F3. 502 Mesilau (Level 5)

Chair: Nik Hakimi Nik Ali (Universiti Teknologi MARA & Shah Alam, Selangor, Malaysia)

11:30 Optimizing Microgrid Profits Through Automated P2P Energy Trading Using Blockchain and BESS

Sajarupan Tharumaraja (University of Sri Jeyawardenapura, Sri Lanka & University of Jaffna, Sri Lanka); P. L. M Prabhani and Akila Wijethunge (University of Sri Jayewardenepura, Sri Lanka); Janaka Ekanayake (University of Peradeniya, Sri Lanka)

The integration of Distributed Energy Resources (DERs), such as rooftop photovoltaic (PV) systems and Battery Energy Storage Systems (BESS), enables peer-to-peer (P2P) energy trading in microgrids, enhancing grid flexibility and optimizing operational management. This study presents an automated, blockchain-enabled framework for very short-term (VST) P2P trading, tested using Sri Lanka's tariff data to harness the economic and operational potential of decentralized energy systems. The Intelligent Prosumer Energy Node (IPEN) facilitates autonomous energy trading through real-time monitoring, VST demand forecasting, recommendations from the OpenDSS Demand-Side Management (O-DSM), and user-guided decisions. Similarly, the Intelligent Consumer Energy Node (ICEN) autonomously executes trading based on power demand monitoring, forecasting, and O-DSM guidance. The blockchain network, built on Hyperledger Fabric, secures and transparently manages transactions across five organizations, supported by multiple channels and smart contracts. Three trading models, Feed-in Tariff (FiT), P2P without storage, and P2P with BESS, were evaluated across prosumer-to-consumer ratios of 25:75, 50:50, and 75:25. Results show that automated P2P trading outperforms FiT, with BESS providing the highest economic gains. Prosumers achieved up to 35.1% higher profits, while consumers reduced costs by up to 15.2%, demonstrating the system's potential for scalable microgrid deployment.

11:45 Innovative Solutions for Smart Grids: Direct Ammonia Fuel Cells and Smoothing Filters for Solar and Wind Power Stabilization

Ahmed Intekhab Rohan and Tasfia Akter Ridita (Islamic University of Technology, Gazipur, Bangladesh); Hasanur Zaman Anonto (American International University-Bangladesh, Bangladesh); Md Ismail Hossain (American International University Bangladesh, Bangladesh); Anup Kumar Roy and Sudipto Roy (Lamar University, USA); Riadul Islam and Abu Shufian (American International University-Bangladesh, Bangladesh)

The paper proposes a combined approach of Solar and wind power fluctuation equalization through Electrochemical Ammonia Synthesis (EAS) and Direct Ammonia Fuel Cells (DAFCs), supplemented by different smoothing filters. When full of renewable energy, the surplus is turned into ammonia and stored and later turned back into electricity when there is a shortage. It uses Moving Average, Moving Median, Moving Regression, Gaussian and Savitzky Golay filters on real wind and solar profiles in MATLAB and uses EAS/DAFC models in Engineering Equation Solver (EES). Findings indicate that the MR filter provides the best compromise between noise attenuation and trend fidelity and peak capacity demands of ammonia production and fuel cell output are lowered by approximately 10-20% relative to raw profiles. SG and Gaussian filters have nearly similar advantages, whereas MA and MM are suboptimal because of either lag or inconsistent trend. This is the novelty of the work because a comparative evaluation of advanced smoothing methods in an ammonia-based storage cycle has never been carried out, and results are used to inform actionable recommendations regarding storage cycle sizing and cost minimization. The suggested methodology increases the stability of the grid, reduces the sizing of subsystems, and enables the cost-efficient implementation of ammonia fuel cells to integrate renewable energies.

12:00 Towards Sustainable Smart Cities: IoT and Big Data-Driven Anomaly Detection in Building Management

Irin Laila Parvin (Bangladesh University of Engineering and Technology, Bangladesh); Md. Arif Rahman (Ahsanullah University of Science and Technology, Bangladesh); Tahsina Farah Sanam (Bangladesh University of Engineering and Technology, Bangladesh)

For the development of sustainable smart cities, the combination of Internet of Things (IoT) and Big Data Analytics plays a vital role while IoT constitutes interconnected sensors to generate a data-rich environment, enabling real-time measurements and contextual information for smart buildings with significant benefits and challenges in security, data acquisition, processing, and storage. To identify anomalous devices in smart buildings, there are various existing statistical and classical Machine Learning models. But they are significantly inefficient to deal with non-linearity with average reported performance. Comprising the LSTM Autoencoder model and implementing advanced data analysis and machine learning techniques, this study represents recent research on anomaly detection with the unique CU-BEMS data set, which offers valuable insights for research applications. This technique includes various data-intensive approaches, along with lightweight methods tailored for edge and in-node computing. The paper highlights state-of-the-art methods for detecting anomalies in sensor systems by addressing challenges such as sensor miniaturization, energy efficiency, security, and data heterogeneity.

12:15 Occupancy Driven Optimization for Comfort and Energy Management in Smart Building

Ghulam Fizza (Universiti Kuala Lumpur, British Malaysian Institute, Malaysia); Kushsairy Kadir (Universiti Kuala Lumpur British Malaysian Institute, Malaysia); Haidawati Nasir (Universiti Kuala Lumpur, Malaysia); Mohammad Rashid (Universiti Teknologi Malaysia (UTM), Malaysia)

Achieving an optimal trade off between comfort and energy consumption in smart buildings remains a complex challenge, especially under dynamic occupancy conditions. Traditional optimization frameworks often neglect occupancy variation, resulting in energy inefficiencies or compromised occupant comfort. This study presents an occupancy aware optimization framework in which the objective weight between Comfort Index (CI) and Energy Gain (EG) remains fixed, while environmental constraints are dynamically adjusted based on occupancy. The approach is tested using three metaheuristic algorithms: Harris Hawks Optimization (HHO), Slime Mould Algorithm (SMA), and Arithmetic Optimization Algorithm (AOA), on a real world smart office dataset comprising 528 hourly observations. Results show that HHO achieved the highest average CI during occupancy (0.9341) with comparatively low energy consumption (1327 kWh), whereas SMA delivered the highest CI during non occupied periods (0.7195) with an average energy consumption of 1311 kWh. These findings validate the effectiveness of incorporating occupancy signals into smart building control, improving comfort satisfaction and energy efficiency without adjusting the objective weight.

12:30 Integrating Future Load Profile Nominations into Predictive Estimation of System Inertia in Distributed Power Networks: a Review

Ryan D Abella and Abdul Aziz G. Mabaning (Mindanao State University - Iligan Institute of Technology, Philippines)

Power system is undergoing a transition toward renewable energy sources (RES). While the high penetration of RES at the distribution level offers significant benefits, it also introduces challenges-particularly the reduction of rotational inertia, which poses serious risks to frequency stability and grid reliability. To ensure stable operation, it is essential to analyze and accurately estimate system inertia to maintain frequency stability. This paper reviews the emerging concept of an enhanced predictive estimation model that integrates Future Load Profile Nominations (FLPN). FLPN provides utilities with improved foresight into upcoming demand changes by enabling early identification of periods and locations that may experience high inertia stress. This review consolidates current research on system inertia and explores how FLPN can improve existing demand estimation practices to support better coordination in distributed energy networks. Aligned with the United Nations's Sustainable Development Goals (SDG) 7 (Affordable and Clean Energy) and 13 (Climate Action), the findings suggest that FLPN holds potential as a foundation for future developments in control strategies, virtual inertia planning, and the broader goal of establishing resilient, low-inertia power systems.

12:45 Collaborative Battery Scheduling of Microgrids: N -Agent Reinforcement Learning Approach

Alpha M M (APJ Abdul Kalam Technological University, India & College of Engineering Trivandrum, India); Hari Kumar R and Lal Priya P S (College of Engineering Trivandrum, India)

This paper proposes a collaborative battery scheduling framework for interconnected microgrids using an (n)-agent reinforcement learning (RL) approach, with the primary objective of minimizing the net cost to the main grid. Each microgrid is modeled as an autonomous agent equipped with local photovoltaic (PV) generation, load demand, and a battery energy storage system (BESS) . The agents learn optimal charge-discharge strategies using multi-agent Q-learning while coordinating energy sharing with neighboring microgrids when beneficial. In order to analyze how well the proposed approach performs, simulation studies are conducted for two scenarios: independent microgrid operation and interconnected cooperative scheduling. A custom Markov Decision Process (MDP) environment is designed to incorporate battery dynamics, power balance and shared actions. The proposed methods are trained and tested using real world load and solar profiles. The results demonstrate that interconnection and collaboration among agents lead to significant reductions in grid dependency and overall cost, compared to isolated operation. The framework ensures efficient battery utilization and promotes distributed energy cooperation in smart grid environments.

Electronics, Circuits & Devices (ECD)

Track B5F4 ECD 5: Electronics, Circuits & Devices (ECD) 5

Room: F4. 503 Dinawan

Chair: Zuhaina Zakaria (Universiti Teknologi MARA, Malaysia)

11:30 Transparent Wideband Fractal Antenna for Modern Communication Systems Using Screen-Printed Silver Nanoparticles

Khaloud Mohammed Nasser AlJahwari (Chung-Ang University, Korea (South)); Abdullah Abdullah (University of Oulu, Finland); Hamza Ahmad (Universiti Teknologi Malaysia, Malaysia); Fauziahanim Che Seman and Ayesha Ayub (Universiti Tun Hussein Onn Malaysia, Malaysia); Nida Nasir (NED University of Engineering and Technology, Pakistan)

This work presents a transparent, wideband fractal antenna which can be fabricated using silver nanoparticles (AgNPs) and screen-printing technology. The antenna features an octagonal-shaped mesh monopole patch and a mesh ground plane, both printed on a transparent polyethylene terephthalate (PET) substrate. The proposed antenna has a compact size of 10 × 11 × 0.55 mm³. The antenna operates over a wide frequency range, from 7.6 to 20.5 GHz, covering both the X-band and Ku-band, and partially overlapping K-band with a total bandwidth of 12.9 GHz. The proposed antenna achieves a peak simulated realized gain of 1.97 dBi, a radiation efficiency of 93%, and a fractional bandwidth of 91.8 %. The use of flexible and optically transparent PET substrate enables deployment on curved or see-through surfaces. With the combination of compact size, wideband performance, cost-effective fabrication, and optical transparency, the antenna shows strong potential for use in radar systems, Satellite communications, and some military and aerospace applications.

11:45 ECG P-QRS-T Wave Peak Based Interval Variability for Interpretable Coronary Artery Disease Screening

Dhaladhuli Jahnavi and Ashutosh Dash (Indian Institute of Technology Kharagpur, India); Kayapanda Mandana (Fortis Healthcare Limited, India); Sundeep Khandelwal and Aniruddha Sinha (Tata Consultancy Services, India); Nirmalya Ghosh (Indian Institute of Technology Kharagpur, India); Amit Patra (Indian Institute of Technology, Kharagpur, India)

There is growing interest in developing non-invasive, widely accessible screening approaches for coronary artery disease (CAD) that are compatible with wearable technology. Prior studies have examined the association of CAD with abnormalities in the QRS-T and PR segments of the electrocardiogram (ECG), which require accurate identification of the onsets and offsets of P-QRS-T waves. This study proposes an alternative approach that extracts variability features from interval time series defined solely by ECG P-QRS-T wave peaks, thereby overcoming the persistent challenge of precise onset and offset detection. The proposed supervised hybrid feature selection approach identified an optimal feature combination that achieved 92% CAD/Non-CAD classification accuracy on the validation dataset. It surpassed popular feature selection algorithms, including LASSO, mRMR, BBA, BCS, and BGW. On a blind test dataset, the selected features enabled the final trained ensemble classifier to achieve 88% accuracy, correctly identifying 86% of subjects with CAD and 92% of subjects without CAD-outperforming state-of-the-art methods evaluated on the same dataset. Furthermore, validation on a manually annotated QT database revealed notable correlations between proposed features extracted from peak based and clinically relevant interval time series, supporting their potential interchangeability when automated onset-offset detection is unreliable.

12:00 Design of a Capacitive-Based PFM DC-DC Converter with Adaptive Stage Control for WSN Applications in 65nm CMOS Technology

Gene Fe P Palencia (Mindanao State University - Iligan Institute of Technology, Philippines); Abdulbasit Gamoranao and Nieva Mapula (MSU-Iligan Institute of Technology, Philippines)

This paper presents the design and simulation of a capacitive-based Pulse Frequency Modulation (PFM) DC-DC converter with adaptive stage control, implemented in 65nm CMOS technology. Targeted for low-power energy harvesting in Wireless Sensor Networks (WSNs), the proposed architecture mitigates limitations of conventional charge pump (CP) designs, including threshold voltage losses, reversion loss, and inefficient power conversion efficiency (PCE) under varying loads. The converter integrates a high-efficiency Modified Cross-Coupled Charge Pump (MCCCP) and an adaptive stage controller using a hysteresis comparator to dynamically adjust the voltage conversion ratio (VCR) based on input voltage. Post-layout simulation results show peak PCEs of 88.23% in Doubler mode and 81.47% in Tripler mode, with consistent voltage regulation across a 0.45 V to 0.7 V input range. Compared to existing solutions, the design achieves better output power, efficiency, and adaptive performance. These attributes make the converter highly suitable for ambient energy harvesting applications in low-power Internet of Things (IoT) nodes.

12:15 UVB Generation Using TIR-Based ORQPM Technique Stimulated by Faraday Rotation

Moumita Saha and Boilla Srinu (VIT-AP University, India)

The analysis numerically demonstrates the generation of 320 nm, a continuous-wave, ultra violate B (UVB), second harmonic using a thin film coated magneto optic (MO) crystal. The considered propagation manner is total internal reflection (TIR). The adopted phase-matching technique is optical rotation quasi phase matching (ORQPM), which significantly enhances second harmonic generation (SHG) efficiency. The required polarization rotation has been made possible by applying an adequate magnetic field. The applied magnetic field triggers the Faraday rotation inside the MO crystal. The thin film, has been utilized to modulate the phase-shifts due to the p- and s-polarized light as they propagate through the interface between slab and film at the time of TIR. A peak conversion efficiency of 22 % has been attained by a computer aided simulation. In the UVB spectrum, it signifies a high performance. The analysis accounts for surface roughness, absorption losses, and nonlinear law of reflection, resulting in near realistic simulation results. The proposed technique provides a controllable way for producing UVB radiation, with potential uses in ultrafast spectroscopy and dermatological phototherapy.

12:30 Comparative Analysis of Marker Based and Marker Less Motion Capture Systems for Shoulder Joint Angle Prediction

Deepa Sri R (Anna University, India & Sri Sivasubramaniya Nadar College of Engineering, India); Pravin Kumar (SSN College of Engineering, India); Kavitha A (Professor & HoD, India); S Saranya (Sri Sivasubramaniya Nadar College of Engineering, India)

Marker-based motion capture system (MCS) is the gold standard for analyzing human motion. But in the real world its usage in large-scale applications is limited by its inherent inaccuracy and practical difficulties. These real-world difficulties can be addressed with a marker-less motion capture system. However, its accuracy in quantifying joint Range of Motion (ROM) has not been verified across shoulder movements. In this study, we simultaneously captured marker less and marker-based motion data on six healthy participants performing shoulder abduction and adduction movements. Based on the results we calculated the RMSE value for abduction 0.80° and for adduction 0.683° between both the systems during each movement. The findings imply that the accuracy of the markerless Inertial Measurement Unit (IMU) system in ROM measurement is comparable to that of the marker-based system. However, particularly when it comes to shoulder ROM measurements, the direction and positioning of the reflective markers of the OptiTrack system have a significant impact on measurement accuracy.

12:45 Dual Band Wearable Antennas Designed on Day to Day Used Jeans for IEEE 802.11 Network

Subhrashil Nanda (India); Rajendra Prosad Ghosh (Vidyasagar University, India)

In recent days, due to a large-scale increase in healthcare activities, the use of various health monitoring devices and the transfer of data from these devices through Body Area Network (BAN) are rapidly increasing. The antenna used in BAN is a wearable antenna, where wearable fabrics are used to design the antenna. Research reported so far has used low-loss engineered wearable fabric to design antennas. In our work, attempts are made to design antennas on day-to-day used materials. The dielectric constant and loss tangent of a used jeans' fabric have been characterised by using an RF impedance analyzer and a dielectric probe kit that uses Open Open-Ended Coaxial Probe (OECP) technique. The dielectric loss is very high (loss tangent equals 0.1152) with a dielectric constant of 1.7976. Two dual-band antennas operating in the IEEE 802.11 designated Wi-Fi bands are reported here. The antenna height is optimized to get higher radiation efficiency. The maximum radiation efficiency achieved (Antenna-2) is 21% in both bands. The antennas are simulated in CST Microwave Studio, and the antenna with maximum radiation efficiency is fabricated for experimental verification.

1:00 Design and Optimization of Complex Quantum Circuits Targeting near-Term Quantum Processors Using Custom Algorithms and Qiskit Transpiler

Kazi Redwan, Mustakim Ahmed, Md. Faruk Abdullah Al Sohan and Sajedul Islam (American International University-Bangladesh, Bangladesh); Birbal Tamang and Rasmila Lama (Lamar University, USA); Ruja Shrestha (Islington College, Nepal); Abu Shufian (American International University-Bangladesh, Bangladesh)

Quantum computing faces challenges such as noise, short coherence time, and limited qubit connections. These challenges worsen as quantum circuits become more complex. One major issue is the increasing depth of quantum circuits. This research proposes an optimization framework targeting depth and gate count reduction in quantum circuits, specifically for Noisy Intermediate-Scale Quantum (NISQ) devices. The proposed approach combines unitary merging of single-qubit gates, CNOT cancellation, gate commuting, and rotation gate rewriting strategies. Consecutive gates acting on the same qubit, such as R x(θ 1) · R x(θ 2), are algebraically merged into a single rotation, while pairs of redundant CNOT gates are eliminated based on gate identity relations. The technique is implemented using Qiskit and validated across five diverse circuits including complex, random, and multi-qubit configurations. Experimental results show an average depth reduction of 33.33% and gate count reduction of 32.14%, with runtime improvement of up to 25%. For instance, an input circuit with a depth of 7 and gate count of 11 was reduced to a depth of 2 and 4 gates. All optimized circuits preserve functional correctness with a fidelity F ≥ 0.99. High-resolution circuit diagrams are presented to visually demonstrate improvements before and after optimization. Additionally, global phase shifts such as e iπ/4 are preserved or analytically characterized where relevant. This work enhances the viability of quantum computations on near-term hardware and opens pathways for future AI-driven quantum optimizations.

Communication Systems (CS)

Track B5F5 CS 5: Communication Systems (CS) 5

Room: F5. 504 Madai (Level 5)

Chair: Aroland Kiring (Universiti Malaysia Sabah, Malaysia)

11:30 12 Gbaud Visible Light Coherent Communication Based on RRC Pulse Shaped BPSK Modulation and Simplified Coherent Detection Scheme

Zhilan Lu, Fujie Li, Jifan Cai, Fang Dong and Zengyi Xu (Fudan University, China); Chao Shen (Fudan University, USA); Junwen Zhang and Nan Chi (Fudan University, China)

The advancement of current data-intensive services imposes increasingly demands on next-generation mobile communication systems. Visible light communication (VLC) offers significant potential for the development of next-generation communication networks. It has an abundant spectrum resource of up to 400 THz, exhibits immunity to electromagnetic interference, and its blue-green band coincides with the underwater transmission window, enabling high-speed communication with high signal-to-noise ratio (SNR). In this paper, we demonstrated a 532 nm visible light coherent communication (VLCC) system based on a lithium niobate phase modulator, and root raised-cosine (RRC) pulse shaped binary phase-shift keying (BPSK) and coherent detection scheme. Compared to intensity modulation with direct detection schemes, our system mitigates signal distortion induced by frequency chirp and relaxes the demand for high sensitivity at the receiver photodetector. Finally, we successfully achieved 12 Gbaud BPSK signal transmission. To the best of our knowledge, this represents the highest reported data rate for blue-green band visible light coherent communication to date.

11:45 Performance of Active and Passive RIS with Co-Channel Interferers in Cellular System

Nur Adriana binti Mawan Iryawan (Multimedia University, Malaysia); Azwan Mahmud (Multimedia University & Telekom Malaysia, Malaysia); Azlan Abdul Aziz (Multimedia University, Melaka, Malaysia); Syamsuri Yaakob (Universiti Putra Malaysia, Malaysia); Nor Azhar Mohd Arif (ELMU, Malaysia)

This paper presents performance analysis and comparison of active and passive Reconfigurable Intelligent Surface (RIS)-aided wireless systems in cellular system with co-channel interferers under realistic channel conditions. A simple and efficient analytical framework based on the Moment Generating Function (MGF) approach is developed to evaluate the downlink ergodic capacity and energy efficiency, considering Nakagami-m fading, path loss, and co-channel interference (CCI) from neighbouring base stations. In the passive RIS configuration, the reflected signals are phase-shifted without amplification, limiting the received signal strength under severe path loss. In contrast, the active RIS employs signal amplification at each element, enhancing the received power but introducing additional amplifier noise and increased circuit power consumption. Closed-form expressions for both ergodic capacity and energy efficiency are derived and validated against Monte Carlo simulations, showing a close match with theoretical predictions. The results reveal that the active RIS-aided system consistently outperforms the passive counterpart in terms of capacity and energy efficiency across various scenarios, such as the number of RIS elements, path loss exponent, amplifier gain, and interference probability, despite its higher power consumption.

12:00 BLSQ: AI-Enhanced Performance Framework for Wireless Multihop Networks

Zhihan Cui (Japan Advanced Institute of Science and Technology, Japan); Yuto Lim (Japan Advanced Institute of Science and Technology (JAIST), Japan); Yasuo Tan (Japan Advanced Institute of Science and Technology & National Institute of Information and Communications Technology, Japan)

Multi-server wireless multihop networks (MWMNs) are critical for modern communication systems, enabling efficient data transmission between devices and servers. However, the complexity of determining optimal server selection and multihop path planning in such networks often results in high interference, high network latency, low network capacity, and reduced network performance. To address these challenges, this paper proposes a two-stage network optimization scheme for MWMNs, using Broad Learning System and Q-learning, called BLSQ. First, the Broad Learning System (BLS) is employed to allocate servers to devices based on their location and computational requirements. Second, a Q-learning algorithm is introduced to optimize multihop path selection, aiming to maximize network capacity while minimizing interference. The proposed approach is evaluated based on different path selection methods in extensive simulations. Results demonstrate that our method significantly reduces network interference, increases network capacity, and achieves lower transmission time, providing a possible approach for optimizing wireless in MWMNs.

12:15 Reconfigurable Intelligent Surface-Aided Spatial Modulation with Signature Constellation

Fanyu Zeng (Macao); Yuyang Peng, Qi Jin, Ming Yue and Runlong Ye (Macau University of Science and Technology, Macao); Liping Xiong (Dongguan Polytechnic, China)

In recent years, a number of new technologies have emerged in the sixth generation (6G) wireless communication area. Reconfigurable intelligent surface (RIS) technologies have been widely studied because of its strong flexibility for signal adjustment. RIS is composed of multiple reflective units which are capable of adjusting their radiation characteristics. By dynamically adjusting the phases of the units, RIS can precisely reflect signals to the receiver and improve the quality of received signals over current transmission systems Combined with transmit spatial modulation (TSM), the RIS-aided TSM (RIS-TSM) system can significantly boost spectral efficiency (SE) while enhancing the received signal quality. In this paper, in order to overcome the effect of correlation among transmit antennas in RIS-TSM systems, we propose an RIS-aided transmit signature constellation based spatial modulation (RIS-TSSM) scheme with a complete system model and expressions. Simulation results depict that RIS-TSSM system can achieve better performance than the RIS-TSM system in the presence of antenna correlation.

12:30 Uniformity Tests on Image Steganography Based on Syndrome-Trellis Codes Without Stego-Keys

Hoover H. F. Yin (The Chinese University of Hong Kong, Hong Kong)

Images are a common type of digital media in computer networks, appearing in web pages, instant messaging, cloud storage, etc. Minor distortion of an image is probably not detectable, so it is possible to hide secret messages inside images sent through the network, thus becoming a potential vulnerability. In image steganography, most steganalysis tools focus on classifying whether each input image is a stego image or not. The algorithm for extracting/decoding secret messages is not used by these tools, although we can assume the knowledge of this information under Kerckhoffs's principle. To adhere to Kerckhoffs's principle, one-time stego-key can be used, but key exchange is challenging in scenarios that apply steganography. We consider steganography based on syndrome-trellis codes (STC) without stego-keys, where STC is a powerful embedding scheme that can minimize the embedding distortion and handle wet pixels without extra effort. As secret messages are usually considered as uniformly random strings, we investigate whether the decoded strings from normal images are also uniform, i.e., statistical detectability. By applying the NIST SP 800-22 Rev. 1a statistical test suite, we show that these decoded strings are instead highly non-uniform, unless the pixels are randomly shuffled. We also demonstrate that the non-uniformity may be applied for pooled steganalysis in theory, even when the extraction includes random shuffling. This suggests that minimizing distortion is not the only metric for measuring security.

12:45 From Noise to Clarity: Emerging Trends in Speech Enhancement for Real-Time Communication

Preethi Sunke and Senthil Mani (Google, India)

Real-time communication demands speech that is both intelligible and natural, even in noisy environments. This paper traces the progression of speech enhancement from traditional DSP techniques-such as spectral subtraction and Wiener filtering-to modern deep learning and diffusion-based models. While classical methods offer low latency and interpretability, they falter in complex, non-stationary noise. Deep neural networks, including CNNs, RNNs, and transformers, brought adaptive, data-driven noise suppression with superior performance. Most recently, diffusion models have redefined the state-of-the-art, enabling high-fidelity speech reconstruction from heavily corrupted inputs. We present a comparative analysis of these approaches in terms of effectiveness, latency, and deployment feasibility, and highlight the promise of hybrid models that unify DSP precision with generative AI power.

Most recently, diffusion models have redefined the state-of-the-art, enabling high-fidelity speech reconstruction from heavily corrupted inputs. We present a comparative analysis of these approaches in terms of effectiveness, latency, and deployment feasibility, and highlight the promise of hybrid models that unify DSP precision with generative AI power.

Computing & Computational Intelligence (CCI) 2

Track B5F6 CCI 5.3: Computing & Computational Intelligence (CCI) 5.3

Room: F6. 505 Sepilok (Level 5)

Chair: Raja Jamilah Raja Yusof (Universiti Malaya, Malaysia)

11:30 Mobile Application for Enhance Sustainable Tea Farming in Sri Lanka

Shashika Lakmini Lokuliyana (Sri Lanka Institute of Information Technology, Sri Lanka); Pipuni Wijesiri (University of Moratuwa, Sri Lanka & Sri Lanka Institute of Information Technology, Sri Lanka); Sahan Chamuditha Kulathunga, Navodi Perera and Moksha Koongahage (Sri Lanka Institute of Information Technology, Sri Lanka)

Ensuring sustainable tea farming requires intensive monitoring of plant conditions, nutritional status, and disease infections. To that end, this research presents a smartphone application that is powered by machine learning to assist Sri Lankan tea farmers in identifying fertilizer and chemical deficiencies, predicting tea yield quality, and detecting diseases at early stages. The system makes use of a trained machine-learning model to scan images of leaves for relevant characteristics to provide instant feedback through a user-friendly smartphone interface. The app offers advice to farmers to improve yield and reduce crop loss. This approach enhances accuracy in farming, minimizes reliance on over-fertilization, and assists in efficient farming methods. The given system is designed to target small scale and far-away farmers to make it more popular in diversified agricultural lands. The research involves mass-scale agricultural image dataset collection and processing, deep learning model training, and deployment of a robust mobile application for field implementation. Outputs strive to contribute to Sri Lankan smart agriculture by allowing farmers to make data-driven decisions to ultimately improve productivity and sustainability.

11:45 A Machine Learning Framework for Data-Scarce Regression Using SMOGN with Joint Hyperparameter Optimization: A Case Study with Cricket Performance Prediction

Harthik Manichandra Vanumu (Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal, India); Paranjay Lokesh Chaudhary (Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal, India.); Usha Moorthy (Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal, India); Syed Anwar Ali (Manipal Academy of Higher Education, Bengaluru, India)

This study presents a machine learning framework to improve predictive accuracy in regression under data scarcity, a prevalent challenge in predictive modeling. A key contribution is a joint hyperparameter optimization strategy that integrates data augmentation with model training, outperforming traditional sequential approaches. Our approach simultaneously tunes SMOGN (Synthetic Minority Over-Sampling Technique for Regression with Gaussian Noise) and multiple regression models using Optuna, identifying optimal parameter combinations that sequential methods may miss. The framework was evaluated using k-fold cross-validation and multi-seed experiments on season-long batting performance prediction in the Women's Premier League (WPL), an emerging cricket league with limited historical data. The results show that tree-based ensembles consistently outperform linear models, with top performers achieving mean test R 2 values above 0.89. CatBoost achieved the highest mean test R 2 of 0.9075, with a standard deviation of 0.0162. An ablation study without SMOGN confirms the importance of this integrated augmentation strategy. By treating the pipeline as a fully integrated system, the framework provides a practical approach to predictive modeling under severe data constraints, with applicability across domains such as sports analytics, finance, and healthcare, and can serve as a blueprint for enhancing data synthesis within AutoML pipelines.

12:00 Enhanced Phishing Payload Detection Using Fine-Tuned DistilBERT and XAI-Based NLP Models

Sourav Datto, Delower Hossen Tuhin, Mustakim Ahmed, Kazi Redwan, Md. Faruk Abdullah Al Sohan and Abu Shufian (American International University-Bangladesh, Bangladesh); Birbal Tamang and Rasmila Lama (Lamar University, USA); Ruja Shrestha (Islington College, Nepal)

Phishing attacks are a major cybersecurity concern. These attacks continue to grow in complexity and often bypass traditional detection systems by imitating legitimate communication payloads. Many existing models, especially classical machine learning techniques, lack the ability to detect hidden or adversarial phishing payloads. They also offer limited transparency in their predictions. This research presents a phishing payload detection approach using a fine-tuned DistilBERT model. The methodology includes dataset preprocessing, model fine-tuning, adversarial training, explainability analysis, and performance evaluation. DistilBERT, a lightweight transformer model, is fine-tuned to detect phishing payloads with improved accuracy and robustness. Adversarial training is applied to defend against input manipulation. Explainable AI (XAI) techniques such as LIME and SHAP are used to interpret the model's predictions. This research shows that DistilBERT achieves a classification accuracy of 98.52% and an AUC score of 0.9993, outperforming traditional machine learning models. It also maintains low false positives and high recall. This research improves the reliability of phishing detection and provides interpretable outputs for security analysis. The results demonstrate that the proposed framework strengthens phishing detection strategies and increases resilience to adversarial attacks. The results are based on a single publicly available phishing email dataset and further validation across diverse datasets and real-world environments is required, with the scope of the findings limited to email-based phishing detection.

12:15 Efficient SpMV for GPUs Using Variable-Sized Vectors and CSR Restructuring

Vishal Manjibhai Pateliya (IIIT Allahabad, India); Anshu S Anand (Indian Institute of Information Technology Allahabad, India)

Sparse matrix-vector multiplication (SpMV) is crucial in many engineering applications, including machine learning, data analytics, and numerical simulation. As matrix sizes expand exponentially, efficient parallel algorithms become increasingly crucial. This work reviews some of the well-known parallel computing strategies, exploring their effectiveness in enhancing algorithmic performance. We focus on techniques for load balancing and data distribution in parallel SpMV, aiming to maximize computational resource utilization. We propose two algorithms for SpMV that attempt to overcome the shortcomings of Flat-SpMV, namely the use of fixed-size vectors and the performance degradation due to irregular matrices. In the first work, we use variable-sized vectors to reduce the partial results, as opposed to fixed-sized vectors used in Flat-SpMV. Further, as the irregular nature of the matrix caused Flat-SpMV to perform slower in warp-level analysis, we developed a new method by reorganizing the sparse matrix structure and applying a variant of Flat-SpMV to it. It outperformed other approaches on more than 70% of the input matrices.

12:30 A Multilingual Intelligent Document Processing System

Ravi Kishore Kodali (National Institute of Technology, Warangal, India); Varsha Sanga (CloudAngles, India); Sai Veerendra prasad Kuruguti (National Institute of Technology Warangal, India); Lakshmi Boppana (National Institute of Technology, Warangal, India)

In today's digital world, processing multilingual documents is critical for business, legal tasks, and information retrieval. This study describes a Multilingual Document Processing System that uses Optical Character Recognition (OCR) and Retrieval-Augmented Generation (RAG) to extract, query and summarize text in multiple languages. The system employs advanced OCR models to correctly recognize text from scanned documents, images, and handwriting in various scripts. By incorporating RAG, it improves comprehension and response generation, allowing users to retrieve and summarize information in English even when the original language is different. This approach takes advantage of recent advances in natural language processing, large language models (LLM), and multimodal AI to address challenges in multilingual data accessibility, knowledge synthesis, and real-time communication. The system provides a scalable AI-driven solution to improve document processing, eliminate language barriers, and increase global user engagement. AWS services support scalable document processing but cold starts in AWS Lambda hinder real time tasks.

12:45 AI-Based Visual Inspection System for Oxygen Indicator Detection in Food Packaging

Ari Aharari (Sojo University, Japan); Kosei Oghushi (SOJO University, Japan)

Oxygen indicators are widely used in packaged processed foods to visually signal oxygen levels by changing color-typically from pink (safe) to blue (defective). However, relying on human visual inspection during final quality checks often results in oversight, leading to the unintentional shipment of faulty products and subsequent customer complaints. This research proposes an AI-based defect detection system that automates the inspection process using the YOLOv8 object detection algorithm. The system consists of a real-time visual recognition module integrated with a laptop, USB camera, photoelectric sensor, and shielding unit, enabling precise detection of color changes in oxygen indicators after product packaging. The proposed system successfully classifies oxygen indicator and on-site evaluations demonstrated a significant reduction in undetected defects and improved inspection efficiency. While challenges such as misaligned packaging and occluded indicators remain, the results highlight the potential of AI-powered visual inspection systems to enhance quality assurance, reduce labor dependency, and support smarter manufacturing practices in the food industry.

Engineering Technologies & Society (ETS)

Track B5F7 ETS 5: Engineering Technologies & Society (ETS) 5

Room: F7. 506 Selingan (Level 5)

Chair: Melvin Gan Status (Universiti Malaysia Sabah (UMS), Malaysia)

11:30 Feature Characterization to Aid in Patellofemoral Pain Syndrome Diagnosis

Gabriel Rodnei M Geslani, Edison Roxas, Emmanuel Guevara, Seigfred V. Prado, Paul Desmond C. Ong, Bernard B Graycochea, Nhaya Marella D Antonio, Julian T Lucina, Sean Clarenz C Joson, Warren Denzel F. Cheng, Consuelo B. Gonzalez-Suarez, Jan Tyrone Cabrera, Timothy Nazareno, Emily Rose D Nacpil, Ivan Neil Gomez and Jazzmine Gale S. Flores (University of Santo Tomas, Philippines)

Patellofemoral Pain Syndrome (PFPS) is a condition that causes pain at the front of the knee, particularly affecting active individuals such as athletes and military personnel. Accurate diagnosis remains challenging, as Magnetic resonance imaging (MRI) methods are often costly. Various applications, including preprocessing and segmentation techniques, have assisted clinicians by improving the quality of patellar tendon imaging. However, clinical observations and measurement of PT-TG distance were affected by factors like practitioner probe angles and patient skin thickness. Therefore, this study aimed to aid clinicians by analyzing key features from ultrasound (US) imaging, focusing on textural and morphological characteristics alongside biological markers. Datasets were collected by the Research Center for Health Sciences at the University of Santo Tomas and included twenty-seven (27) participants, fourteen (14) with PFPS and thirteen (13) without PFPS. Fifty-one (51) features were extracted and analyzed through Feature selection techniques, including Statistical Analysis with Principal Component Analysis (PCA), Recursive Feature Elimination (RFE), and Mutual Information in selecting the most optimal features. These features were validated using sensitivity, specificity, recall, and accuracy. Twenty-three (23) features were selected using statistical analysis with PCA, composed of fifteen (15) textural features, four (4) morphological features, and four (4) biological features were selected. The final model achieved an F1 score of 89% for classifying non-PFPS and 81% for classifying people with PFPS, with overall accuracy of 86%. By analyzing these selected features, the study aims to enhance the evaluation of ultrasound images and biological markers for PFPS detection, contributing to better patient care.

11:45 A Machine Vision-Based FSL Tutor with Static and Dynamic Gesture Recognition and Real-Time User Feedback Using MediaPipe Frameworks

Pocholo James Loresco (De La Salle University, Manila & Far Eastern University Institute of Technology, Philippines)

Filipino Sign Language (FSL) is an invaluable tool for communication within the deaf and mute communities, yet there is a shortage of proficient special education teachers and accessible learning materials. Current research on FSL recognition is limited to basic detection, often invasive, and lacks comprehensive systems that provide feedback to users. Additionally, FSL features unique static and dynamic gestures, including contractions, distinct from other sign languages. This study presents the development of a machine vision-based FSL tutor that leverages the MediaPipe framework-specifically, MediaPipe Hands for static gesture recognition and MediaPipe Holistic for full-body dynamic gesture tracking. LSTM networks were used to classify dynamic gestures based on sequential landmark data to capture temporal dependencies in sign execution. The system supports a desktop application platform enabling learners to engage in interactive modules with real-time feedback through visual prompts and audio cues. It utilizes 42 static hand feature landmarks and over 1,662 key points derived from hand, pose, and facial data to ensure accurate recognition and feedback. A total of 50 essential FSL gestures-aligned with the kindergarten curriculum-were modeled, covering alphabet knowledge, vocabulary development, self-introduction, and polite expressions. Performance evaluation using computer vision metrics demonstrated high recognition accuracy for both gesture types. In addition, the System Usability Scale (SUS) and statistical comparisons with traditional instruction methods confirmed the platform's effectiveness and user acceptability. The results validate the system as a comprehensive and accessible solution for FSL education, particularly suited for early learners and self-guided instruction.

12:00 Public Auditability is Not Enough: on the Importance of Separation of Duty in e-Voting Systems

John Ultra (University of the Philippines Diliman, Philippines); Susan Pancho-Festin (U. of the Philippines, Philippines)

Although the e-voting protocol literature is rich, most proposals focus only on the security of certain phases of an election process, such as voting and counting, while assuming the security of other phases. Voting protocols exploit the concept of public auditability to justify the security of the schemes. However, we argue that any irregularity observed later in the election process casts doubt on its integrity. The cost of fixing a mistake or correcting a wrong outcome increases over time. The ability to detect and correct these mistakes, deliberate or not, early on in the election process is critical to maintaining integrity and transparency. To this end, we explore traditional access control concepts such as role-based access control and separation of duty as mechanisms for ensuring the security of elections using e-voting systems. We make two contributions. First, this paper presents the design and implementation of Helios RBAC, an extension of the Helios web-based voting system that supports role-based access control (RBAC) and separation of duty policies. Second, we provide a security analysis of separation of duty policies in e-voting systems, where we examined different threat models. Contrary to intuition, we show that a simple majority is insufficient to ensure an honest majority decision in the presence of imperfect and malicious election administrators.

12:15 Urban Heat Island and Heat Vulnerability Assessment with Ground Validation - a Case Study for Makati City

Charles Andrew C Pasion (Adamson University Philippines, Philippines & National University Philippines, Philippines); Mark Angelo C Purio (Adamson University, Philippines)

Urbanization, characterized by the movement of populations from rural to urban areas, is reshaping societies globally. This transformation is driven by economic, social, and environmental factors that make urban areas appealing, offering better jobs, access to healthcare and education, and improved infrastructure. However, urbanization also presents significant challenges, including the Urban Heat Island (UHI) effect, which exacerbates living conditions in densely populated cities by increasing surface and air temperatures compared to surrounding rural areas. This study addresses the UHI phenomenon and heat vulnerability by conducting an Urban Heat Island and Heat Vulnerability Index (HVI) assessment, incorporating ground validation, for Makati City, Philippines. The UHI Map was developed using Landsat 8 satellite Land Surface Temperature (LST) data. The HVI Map was developed using Landsat 8 Data Product and Demographic data to identify areas at higher risk of adverse heat-related effects. Each indicator scores were normalized, scored, and aggregated to calculate an HVI score for each barangay. Ground validation was performed to verify the results from the UHI and HVI maps. A prototype device was designed, consisting of sensors for ambient temperature and relative humidity, a memory device for data storage, and a battery with a one-month operational lifespan. Three devices were deployed in barangays identified by the HVI Map: two in the most vulnerable barangays and one in the least vulnerable barangay. Over a month, the devices recorded temperature and humidity data, which were analysed to corroborate the maps' findings. The UHI and HVI maps identified areas within Makati City that are most susceptible to heat-related health impacts, offering critical insights for policymakers and urban planners. The integration of socio-economic and environmental factors ensures a holistic understanding of heat vulnerability, enabling evidence-based strategies to mitigate risks. The findings support urban planning efforts aimed at reducing heat exposure through increased vegetation, improved building designs, and the development of heat-resilient infrastructure. Additionally, this research contributes to achieving Sustainable Development Goal (SDG) 11, Sustainable Cities and Communities, by promoting strategies for creating inclusive, safe, and sustainable urban environments.

12:30 A TAM-Guided Mobile Solution to Support Mental Wellness in Higher Education

John Angelo Repollo and John Heland Jasper Ortega (FEU Institute of Technology, Philippines)

Mental health concerns are on the rise among college students in the Philippines, where academic stress and limited access to counseling services continue to pose serious challenges. With mobile technology becoming more integrated into daily life, it offers a practical opportunity to support student well-being through accessible, self-help tools. This study presents the design and evaluation of a mobile application that combines art therapy and sound therapy to help reduce stress and promote relaxation among students in higher education. Guided by the Technology Acceptance Model (TAM), the research explored how users perceived the app's usefulness, ease of use, and overall experience. The app was developed using a blended Agile approach and tested by 50 purposively selected college students experiencing academic stress. Results showed strong user acceptance, with high ratings in ease of use (x̄ = 4.56), satisfaction (x̄ = 4.36), and intention to use (x̄ = 3.96). Perceived usefulness was strongly correlated with both satisfaction (r = 0.73) and continued use (r = 0.78), indicating that the app effectively supported stress relief and user engagement. This study contributes practical insights for integrating mobile wellness solutions in Philippine education, particularly in settings where traditional mental health support remains limited. It encourages the adoption of simple, evidence-based digital tools that promote emotional well-being and help bridge gaps in student support systems.

12:45 FedLEM: Federated Learning with Local Episodic Memory for Data Heterogeneity

Nafas Gul Saadat (Cochin University of Science and Technology & None, India); Santhosh Kumar G (Cochin University of Science and Technology, India)

Federated Learning (FL) is a new learning paradigm that allows collaborative learning among multiple nodes without revealing raw data. However, one important problem faced by the present federated learning algorithms in real-world applications is data heterogeneity. i.e data distribution across clients is different, resulting in low model accuracy, and making the averaged model significantly diverge from the optimal solution. In this study, we propose FedLEM, an efficient federated learning method to address the challenge raised by non-IID data. FedLEM utilizes a small local episodic memory module to selectively store model updates that have been trained on IID batches of data, identified through entropy-based estimation. When a batch exhibits high non-IID characteristics, FedLEM blends the current update with the stored IID updated from memory. This strategy stabilizes optimization and improves model performance under heterogeneous data. By integrating past IID knowledge, this approach helps prevent catastrophic forgetting, improves convergence, and reduces computation and communication overhead in federated learning. FedLEM's effectiveness has been examined on the CIFAR-10 dataset, and the result compared to baseline algorithms: FedAvg, FedProx, FedOpt and Scaffold. FedLEM surpasses baseline algorithms in accuracy and reduces training time by around 16%.

Computing & Computational Intelligence (CCI) 3

Track B5F8 CCI 5.4: Computing & Computational Intelligence (CCI) 5.4

Room: F8. 507 Monsopiad (Level 5)

Chair: Jackel Vui Lung Chew (Universiti Malaysia Sabah, Malaysia)

11:30 Kisan-Mitra: Empowering Farmers with AI-Driven Generative Assistance Transformer Network for Agricultural Advancement

Aditya Oza (IIIT Naya Raipur, India & No, India); Rahul Yadav and Mallikharjuna Rao K (IIIT Naya Raipur, India); Rimjhim Sharma (IIIT NAYA RAIPUR, India)

Agriculture remains the backbone of India's economy, employing over two-thirds of the population and contributing approximately 19% to the national GDP. This paper introduces the Generative Agricultural Transformer (GAT), a novel transformer-based architecture tailored for the Indian agricultural domain. GAT powers Kisan-Mitra, a multilingual generative AI chatbot that delivers context-aware, region-specific guidance aligned with government schemes and agronomic best practices. Unlike generic language models such as BERT or GPT, GAT integrates domain-specific attention mechanisms, cross-lingual embeddings for 12 regional languages, and real-time connectivity with government databases, including the Kisan Call Centre (KCC) and agricultural subsidy portals. As a result, the model is fine-tuned for tasks such as crop advisory, disease diagnosis, seasonal planning, fertilizer optimization, and market price forecasting. GAT achieves a response accuracy of 89.6% and covers over 85% of key agricultural topics, significantly outperforming baseline transformer models. The proposed system is scalable, fully trainable end-to-end, and effectively bridges the digital divide by providing accessible, accurate agricultural assistance to rural communities. This study presents the complete development pipeline of Kisan-Mitra, including dataset construction, architectural innovations, training strategies, and deployment approach. Results indicate that task-specific transformer architectures-when grounded in local context and institutional integration-can substantially enhance the impact and inclusivity of AI-driven agricultural services.

11:45 Benchmarking Hybrid Deep Learning Models for Early Weather Prediction

Deepika Gupta (Indian Institute of Information Technology Vadodara, India); Kushagra Taneja, Chitransh Kumar and Kartik Chugh (IIIT Vadodara International Campus Diu, India)

Weather forecasting is essential for agriculture, disaster management, energy planning, and infrastructure protection. Accurate and timely forecasts support critical decisions, helping to reduce losses and optimize resource use. However, traditional numerical weather prediction (NWP) systems demand high-performance computing resources and long-term historical data, making them impractical for newly deployed or remote weather monitoring stations. To address this limitation, this research proposes and evaluates deep learning-based architectures for weather time series forecasting under varying data availability conditions. Specifically, three standalone models and several novel hybrid models are tested across four temporal scales: 10 years, 5 years, 1 year, and 1 month. The study shows that hybrid models consistently outperform standalone models in terms of accuracy and robustness. Among them, the Temporal Convolutional Network - Long Short-Term Memory (TCN-LSTM) hybrid achieves the best performance, with (R^2) of 0.88 for short-term and 0.99 for long-term temperature forecasting, while remaining computationally efficient. These models offer scalable, accurate forecasting solutions well-suited for early deployment in resource-constrained and data-scarce environments.

12:00 Depth-Based Volume Estimation of Filipino Food Items Through YOLOv8 Custom Dataset

Klarisse Anne C. Mañibo, Francisco G. Joaquim Da Silva, Marie Faye S. Palsimon, Mark Angelo C Purio and Evelyn Q. Raguindin (Adamson University, Philippines)

Accurate food intake tracking is key to a healthy lifestyle, especially for managing diet-related illnesses. Traditional self-reporting methods are often unreliable due to memory and portion size estimation errors. This study introduces a new system that uses object detection and depth mapping to measure the volume of Filipino food items. The system uses a custom dataset with images of three food classes taken from top and side views, each annotated with bounding boxes. It employs the YOLOv8 model for accurate object detection and depth estimation to create 3D food models for precise volume calculation. Achieving 95.4% detection accuracy and a 7.8% average volume error, the system shows promise for dietary monitoring applications. Experimental results prove the system's ability to process the custom dataset effectively, achieving accurate object detection and volume estimations. Through the integration of innovative computer vision methods with culturally appropriate food data, this solution provides a promising method for boosting dietary monitoring and nutritional evaluation in the Filipino community setting. While the model performed well on the three predefined food classes, misclassification of non-target items (e.g., Biko) highlights the absence of a background or rejection class in the training dataset.

12:15 Context-Aware Scene Text Alignment Classification for Mobile Devices

Selvakumar K (Vellore Institute of Technology Vellore, India & VIT Vellore, India); Ayush Kumar and Vivin Varshan S (Vellore Institute of Technology, India); Saptarshi Manna and Sunil Gangele (Samsung R&D Institute, India)

This paper presents a novel lightweight model for text block alignment detection, specifically optimized for on- device deployment in mobile applications. Accurate alignment classification (left, right or center) is critical for downstream tasks such as mobile-based image translation and OCR post- processing. The proposed architecture employs MobileNetV4 as a backbone for efficient feature extraction, integrated with a Feature Pyramid Network (FPN) to enhance context-aware representation across scales. A novel mask-based feature filtering mechanism is introduced to suppress irrelevant visual content and isolate alignment-specific cues. Subsequently, the proposed regions are fed into a lightweight classification module. To ad- dress the inherent differences in contextual dependencies between single-line and multi-line text blocks, the model employs a hybrid feature representation: standard FPN features are used for single- line blocks, while lightweight pyramidal features are used to construct a Pyramidal feature hierarchy for multi-line blocks. Experimental evaluations demonstrate that the hybrid approach achieves superior accuracy compared to using traditional FPN features alone. Furthermore, the model is benchmarked on multiple mobile hardware platforms using Qualcomm's on-device AI execution APIs and the end-to-end inference latency is <10 ms, validating its practicality for real-time deployment.

12:30 Data Oriented Fairness in Cross-Silo Federated Learning

Tomsy Paul (Cochin University of Science and Technology, India & RIT Kottayam, India); Santhosh Kumar G (Cochin University of Science and Technology, India)

Fairness is an active research area in Federated Learning (FL). And Fairness in Contribution Evaluation is an important type in the current literature. Its significance is increased in Cross-silo setting. However the most important component of the FL system, the data is often ignored while studying it. We propose three fairness algorithms from the perspective of the data owner, leveraging respectively the local sample sizes, the sample distributions and both the sample size and sample distributions, offering a unique perspective on fairness in FL. We also provide the implementation of the algorithms in Decentralized FL, a recent area in Federated Learning. A container based implementation makes the development and deployment of the algorithms easy in academia and industry. The experimental evaluation of the algorithms on a typical Cross-silo FL setting with 16 parties show that all the three algorithms provide good fairness to the data owners based on the quantity of data and/or the quality of data.

12:45 Enhanced YOLO Object Detector for Insulator Defect Detection in Power Line Infrastructure

Seema Choudhary (Academy of Scientific and Innovative Research (AcSIR) & CSIR-Central Electronics Engineering Research Institute, India); Sumeet Saurav (CEERI Pilani, India); Ravi Saini (CSIR Central Electronics Engineering Research Institute, India); Sanjay Singh (CSIR-Central Electronics Engineering Research Institute (CSIR-CEERI), India & Academy of Scientific & Innovative Research (AcSIR), India)

Deep learning has shown remarkable capabilities in automatic defect detection in power line infrastructure, but the scarcity of defect-specific labeled datasets often limits its effectiveness. This work addresses the critical challenge of detecting missing disc insulator defects under data-limited conditions by proposing a data augmentation-enhanced YOLOv12 framework. Starting with only 128 original defect images, we systematically applied geometric augmentations, including multi-angle rotations (10°, 20°, 30°) and spatial shifts (horizontal/vertical shifts of 0.1-0.3) to generate 27 distinct variations per image. This strategy expanded the dataset to 3,456 synthetic samples, enriching defect diversity while preserving realistic defect characteristics. The YOLOv12-based framework was evaluated using 5-fold cross-validation, with parallel GPU training used to fully utilize computational resources and reduce training time. Experimental results demonstrate that the diversity of synthetic data, combined with the advanced detection capabilities of YOLOv12, significantly improves model robustness, achieving a 30-34% increase in mAP over non-augmented training and surpassing existing augmentation-based and improved fault detection methods. This study provides a practical approach to overcome data scarcity and advance reliable defect detection in power line inspection applications.

BACK

HOME

Page updated

Google Sites

Report abuse

DAY 2 (29 Oct 2025) Track 5 (11:30am-1:00pm)

IEEE TENCON 2025 (October 27-30, 2025) SICC Kota Kinabalu, Sabah, MALAYSIA

DAY 2 (29 Oct 2025)
Track 5 (11:30am-1:00pm)

IEEE TENCON 2025 (October 27-30, 2025)
SICC Kota Kinabalu, Sabah, MALAYSIA