Browse Items (16481 total)
Sort by:
-
Blockchain-Based Financial Transparency Models for Leather Export Industries in India
The leather export sector in India is confronted with repeated issues of providing financial transparency, traceability, and trust of the stakeholders because of a fragmented payment system and manual records. To resolve these, a blockchain-based financial transparency model is elaborated based on a distributed ledger, which operates under smart contracts, ensuring immutable records of transactions and automated verification. The model uses a secure financial exchange by using hash-based encryption and a consensusbased validation in order to synchronize export payments between decentralized nodes. The Ethereum-based Hyperledger Fabric was simulated to check the accuracy, latency, and scalability of the model. Experimental results indicate that there is a 31.7 % increase in financial traceability, 24.5 % decrease in processing delay, 18.9 % increase in cost efficiency, and 27.6 % high trust score among existing methods. The proposed framework will provide real-time, non-tampered, and verifiable financial transactions, which will provide a long-term solution to the Indian leather export industry with a pathway to transparent and responsible export management. 2026 IEEE. -
AI-Enabled Brand-to-Generic Medication Recommendation System for Pharmacy: Addressing Consumer Brand-Name Bias
In countries like India, many people choose branded medicines over equally effective generic options, which drives up healthcare costs. This paper introduces a smart recommendation system that can tell whether a medication query is about a brand-name or a generic drug. It then offers more affordable alternatives based on the active ingredients. By using natural language processing and retrieval-augmented generation (RAG) with a detailed medicine database, the system accurately classifies and recommends options. Its conversational interface mimics a real pharmacy interaction, helping users make informed choices while saving money. Tests show the system responds quickly, usually within 7 seconds, and provides accurate answers over 80 % of the time for straightforward queries. Ultimately, this tool addresses the information gaps about generic drugs and branded ones and is easy to use for both consumers and pharmacists. 2026 IEEE. -
Integrating Machine Learning with Financial Risk Modeling for Portfolio Management
Financial markets may be unpredictable and volatile; the ability to perform proper risk forecasting and effectiveness in performing an efficient portfolio is of primary importance when making wise investment choices. The nonlinear trends, and time dependence applied in financial data are usually not captured in conventional predictive models. The research is suggesting a new hybrid architecture LSTXplain that combines with and is afforded capabilities of SHAP, and exogenized with LSTM networks as well as Experimental learning. The aim of this paper, which is entitled Integrating Machine Learning with Financial Risk Modeling to Portfolio Management is to combine sequential learning with interpretability in an attempt to deepen financial risk prediction and portfolio optimization. The model is intended to forecast various measurements of financial risk, such as volatility and Value-at-Risk, and is also likely to establish the causes of each of these estimates. LSTXplain uses historical stock prices, technical features and optionally, sentiment scores designed using financial news to train a robust deep learner. Model outputs are then fed through SHAP that allocates a value of importance of a feature and discover that this allows analysts to know and trust what the model does. In order to compare the framework, Yahoo Finance data was applied, and the findings were compared to the traditional models ARIMA, SVM, Random Forest, and MLP. It has a prediction accuracy of over 98 percent which does not just complement the risk forecasting but enables a portfolio management to act. The analysis is a bridge between the performance of DL and explainable AI in the financial risk prediction. Statistical significance were applied to prove that such improvements are significant, and it is established that results are significant at p<0.05. 2025 IEEE. -
OrthoTrace - Fracture Detection from X-Ray Images Using YOLOv5
Rapid and correct diagnosis of bone fractures is important for appropriate and efficient management in clinical practice; however, traditional X-ray interpretation frequently involves diagnostic error, particularly in low resource environments. This study presents OrthoTrace, a YOLOv5-based deep learning system designed for real-time and lightweight fracture detection. The model was trained on a curated set of annotated X-ray images from the publicly available YOLOv5 dataset, which includes a range of upper and lower limb fractures across diverse anatomical regions regions ( 3,500 - 4,000 images; 70:20:10 train/validation/test split) and evaluated using accuracy, mean Average Precision (mAP), precision, recall, and F1-score. On the held-out test set, OrthoTrace achieved 90% accuracy and an mAP of 0.91, outperforming standard CNN baselines and approaching YOLOv8 performance while requiring significantly fewer computational resources. YOLOv5 can also identify fractures in real time, and draw bounding boxes with confidence scores around the specific areas of the X-ray with fractures. OrthoTrace can also be deployed in local or cloud environments. The novelty of OrthoTrace lies in its incredibly rapid inference, while requiring very little computing power. Although promising, the system is limited by dataset size and lack of external validation. This system holds clinical significance as it can support rapid, reliable fracture detection in low-resource or emergency settings, making real-time deployment feasible on standard hospital hardware. Future work will focus on expanding datasets, and enhancing robustness across various medical scenarios. 2025 IEEE. -
Predictive Analytics for Train Timeliness Using Long Short Term Memory and Machine Learning Techniques
Train delays are one of the most persistent issue faced by Indian Railways and it has still been the issue with all the modern infrastructure and increasing travel demands. Traditional methods mainly depend on historical averages and simple modelling which fail to capture complex patterns in delays caused. This research aims to build machine learning and neural network models to analyse historical data from past train journeys and make predictions for future train journeys. Machine learning models include Decision Trees, XGBoost, Random Forest, Extra Trees and a neural network model LSTM to predict the delay for a particular train on a given day based on the previous running status. The highest accuracy of 94.02% was found using LSTM model and a lowest of 72.65% for Decision Tree Regressor algorithm. 2025 IEEE. -
Optimizing Interpretability in Recommender Systems using a Hybrid Model based on Matrix Factorization and Neural Networks
Recommender systems play a crucial role in the direction of user choices in e-commerce, media, and online services, clearly, there is a trade-off between predictive accuracy and interpretability. In this paper, a new hybrid model that combines Matrix Factorization and a Neural Network framework to maximize the performance of recommendation as well as explainability has been suggested. The model uses Latent factor representation of Matrix Factorization to provide the global user item interactions, and the Neural Network component finds nonlinear interaction and contextual patterns in the data. The hybrid architecture is trained and tested on a Kaggle dataset of 100,000 user-item interactions with several numerical and categorical characteristics. It compares to standalone methods in that the system is more superior with an accuracy of 94.5, F1-score of 0.945, mean absolute error (MAE) of 0.087 and root mean squared error (RMSE) of 0.112. It is proven by computational analysis to have efficient training convergence and low inference latency, allowing real-time recommendations on Google Colab. The proposed solution bridges the gap between performance and transparency since it can be applied and is credible by being predictive and understandable at the same time. The study has implications in intelligent, explainable and scalable recommenders systems in diverse areas of application. 2025 IEEE. -
FedDiff-Health: A Privacy-Preserving Generative Framework for Collaborative Hospital Readmission Prediction
Hospital readmission prediction encounters three challenges: data siloing across hospitals due to format incompatibilities, stringent privacy constraints, and the rarity of readmission events. We propose P-Fed-Diffusion, the first framework that enables collaboration across hospitals while keeping patients' data private. Our method automatically aligns heterogeneous data schemas without human intervention, using large language models. Then, we apply conditional diffusion models within a federated learning framework to generate synthetic data for rare readmission events. The framework incorporates formal privacy guarantees via differential privacy. We achieve a dramatic improvement over state-of-the-art methods: while the best prior method achieves 2% recall, we achieve 64% recall-32x improvement, meaning that the method finds over 1,000 additional high-risk patients per hospital annually. Our work opens up a new direction for privacy-preserving collaborative AI across hospitals. 2025 IEEE. -
Impact of Macroeconomic Integration in Hybrid GARCH-GRU Volatility Modelling on Nifty Bank
In countries like India, where banking systems are closely tied to macroeconomic swings, being able to forecast volatility is critical for managing financial risk. Sudden changes in interest rates, exchange movements, or growth expectations can unsettle banks much faster than in mature markets. Econometric tools such as the Generalised Autoregressive Conditional Heteroskedasticity (GARCH) model remain popular because they capture volatility clustering well, but they fall short when the data exhibit nonlinear patterns. Neural networks-particularly Gated Recurrent Units (GRUs)-handle time-series dynamics more effectively, though they tend to miss traits specific to financial volatility. In this work, we put forward a hybrid GARCH-GRU framework that blends the traditional strengths of econometric models with the pattern-learning ability of neural networks, while also folding in key macroeconomic indicators. The framework is applied to the Nifty Bank index and draws on daily records spanning March 2010 to December 2022. Altogether, the dataset includes just over three thousand observations, covering more than a decade of varied market conditions. The framework uses a two-step design: conditional volatility from a GJR-GARCH(1,1) model is first estimated and then used as input, along with macroeconomic variables such as repo rates, exchange rates (USD/INR, CNY/INR, EUR/INR), oil prices, and GDP growth, for the GRU network. Our results indicate that the hybrid model performs noticeably better, cutting the Mean Absolute Error by about a quarter. The error falls from 0.000263 in the baseline GARCH model to 0.000199 under the hybrid design. Among the different factors considered, movements in exchange rates and changes in repo rates stand out most strongly, showing how these macroeconomic signals feed directly into risk management for Indian banks. 2025 IEEE. -
A Comparative Study of Unsupervised Models for Anomaly Detection in Maritime AIS Data
The integrity of global maritime trade is increasingly threatened by deceptive practices such as sanctions evasion and illicit trafficking, often facilitated by the manipulation of vessel tracking data from the Automatic Identification System (AIS). While AIS provides a rich source for monitoring vessel behavior, the vast scale of the data and the novelty of anomalous patterns necessitate advanced, automated detection methods. This paper presents a comprehensive benchmarking study of four dis-tinct unsupervised machine learning architectures for detecting anomalies in historical AIS vessel trajectories. The evaluated models include a Bidirectional GRU (Bi-GRU) autoencoder, a probabilistic GeoTrackNet with A Contrario detection, a two-level grid representation with Isolation Forest, and a multi-model approach combining spatial-thematic attributes with Isolation Forest. We provide detailed mathematical formulations, algorithmic descriptions, and rigorous comparative analysis of each approach, examining trade-offs between temporal modeling, spa-tial context awareness, feature engineering, and computational complexity. Our benchmarking results on 985,700 AIS messages indicate that spatially-aware models (GeoTrackNet, grid-based methods) demonstrate significantly higher sensitivity (6.76%-10.00% anomaly rates) than purely temporal models (0.20%), but at greater computational cost. This study provides practical guidance for model selection based on operational requirements and proposes future directions toward multimodal architectures integrating trajectory analysis with document-based verification. 2025 IEEE. -
A Hybrid Framework for Detecting Hallucinations in Large Language Model Outputs
With Large Language Models (LLMs) continuously growing, they are on a path to replace the search engines soon. No matter how powerful they get, there is a certain level of uncertainty because they tend to hallucinate. Hallucination here refers to generate factually incorrect data, this can include making up names, generating false links and fabricating stories. This makes it extremely difficult to trust large language models. Existing papers provide solutions which are either not monetarily feasible or lack capabilities to build a robust hallucination detector. This paper aims to build a low resource hallucination detector which combines multiple heuristic signals like semantic similarity, self-consistency, external grounding via Wikipedia, NER overlap, flexible numerical check and a quantized LLM Falcon-7b. This eliminates the need to train the model from scratch. Upon evaluating with an input dataset of 50 questions the model was able to achieve an accuracy of 88%. 2025 IEEE. -
Enhancing Time Series Forecasting in Low-Liquidity Markets Using Generative Adversarial Networks
Financial assets that are low liquidity are very difficult to forecast as they are sparsely traded, their volatility is not regular, and scarce historic evidence exists. This paper will explore the hypothesis of whether in this kind of limited environment, generative models can enhance the effectiveness of forecasting. A dual model framework is constructed which contrasts a normal Long Short Term Memory (LSTM) network with TimeGAN based synthetic data augmentation method in 60-day long-range forecasting of the TRY/USD exchange rate. The methodology consists in the training of an LSTM model on real historical sequences and the improvement with TimeGAN generated synthetic sequences with a maintained temporal structure. It has been shown that TimeGAN has a significant effect on the accuracy of the forecasts, the RMSE decreased to 0.0002 by approximately fifty percent, and the R2 grew to 0.9921 by approximately fifty percent. The results suggest that augmentation through GAN enhances generalization of models in thin and dynamic markets. The most important contributions include implementation of TimeGAN to low-liquidity FX forecasting, the assessment of the effects of synthetic data on forecast accuracy and the empirical benchmark of LSTM and TimeGAN in low-volume finance. 2025 IEEE. -
Comparative Analysis of Noise Generated in BGV Homomorphic Encryption: Lattigo vs FHEgen Parameters
Post-quantum cryptography has emerged as a critical field following advances in quantum computing that threaten classical encryption schemes such as RSA and ECC. Fully Homomorphic Encryption (FHE), particularly the Brakerski-Gentry- Vaikuntanathan (BGV) scheme based on the Ring Learning with Errors (RLWE) problem, provides a promising solution for secure computations on encrypted data. A fundamental challenge in BGV implementations is the growth of noise during homomorphic operations, which must remain below a decryption threshold to ensure correctness. This study presents a comparative analysis of noise generation in BGV implementations using two distinct parameter selection approaches: Lattigo's pre-validated generic parameters and FHEgen's automatically generated application-specific parameters. Through empirical measurements using Lattigo v6.1.1, we evaluated five parameter sets across initial noise after encryption, noise expansion during homomorphic multiplication, and overall noise growth patterns. Our results demonstrate that Lattigo N13 achieves marginally lower post-multiplication noise (0.0587 log2 bits, or 4.15% lower in magnitude), though FHEgen achieves substantially higher verified security (210 bits vs. 50-60 bits). However, Lattigo's range of pre-validated parameters (LogN = 12 to LogN = 15) offers greater flexibility for varying computational depths. We conclude that the choice between parameter selection approaches depends on application requirements: FHEgen is preferable for well-defined computational needs with noise optimization priorities, while Lattigo is advantageous when flexibility and extensive validation are critical. This work provides practical insights for FHE practitioners in selecting parameters that balance security, noise management, and computational efficiency. 2025 IEEE. -
Analyzing the Role of LIME and SHAP in Explainable DoS Attack Detection for IoT Systems
Explainable Artificial Intelligence (XAI) based tools such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are extensively used in various detection and prediction approaches. These tools extract feature importance from the datasets and explain the contribution of the features (feature importance) towards detection /prediction output both locally and globally. In the current study a performance analysis is represented on the behaviour of LIME and SHAP explainability towards Denial-of-Service Attack detection in Internet of Things. There are numerous Black-box models including Machine Learning which show high detection accuracies in such case but the output is not interpretable by the security analyst most of the time. this drawback is overcome by introducing LIME and SHAP interpretability to the output of BlackBox model by analysing feature importance of the attack dataset towards detection accuracy. However, LIME and SHAPE has different behaviour towards model-interpretability. SHAP is powerful in global explanation where LIME works efficiently on local interpretation. We have shown that these two different tools perform on same detection accuracies of DoS attack using Machine learning model. A random forest classifier is first selected with high detection accuracy on a simulated DoS attack dataset and at the output SHAP and LIME are executed for achieving both local and global explainability. The comparison shows how SHAP and LIME show strength and weakness in explaining model's behaviour both locally and globally. 2025 IEEE. -
DeepRetina: Transformer-Enhanced EfficientNet for Retinal Disease Classification
Retinal diseases are a major cause of visual impairment in India, which requires precise and automated diagnosis tools.This paper, introduce a two-phase deep learning architecture for classifying five common retinal ailments: Glaucoma, Normal Fundus, Pathological Myopia, Hypertensive Retinopathy, and Cataract. A Swin Transformer (Swin-T) was fine-tuned on augmented retinal fundus images in the first phase to extract domain-adapted feature representations. The transformer utilize such embeddings for learning a regularized EfficientNet-inspired classifier in the second phase, with mixup augmentation and label smoothing for improving generalizability. Comprehensive experiments conducted on a carefully curated dataset of 643 test images validate that our method attains a test accuracy of 93.93%, with high precision as well as recall across all categories. The suggested pipeline strikes a suitable balance between feature abundance with transformer-based adaptation and resilient classification with EfficientNet, providing a viable tool for automated diagnosis of retinal ailments in practical clinical scenarios. 2026 IEEE. -
Real-Time Safety Monitoring for Construction Sites Using RFID and Visual Recognition Technologies
The integrated automated safety monitoring system for construction sites utilizes RFID, Wi-Fi, and vision-based recognition systems to enhance worker safety and ensure adherence to safety regulations. This system combines sophisticated components such as RFID tags and Raspberry Pi, solenoid locks, servo motors, and PIR sensors to provide an exhaustive solution. RFID technology is applied to assign unique tags to each worker, facilitating accurate tracking and identification The Wi-Fi and visual recognition components improve the system's functionalities, enabling wireless connectivity instantaneous data transmission, and verification of appropriate safety gear application. Solenoid locks and servo motors ensure regulated access to hazardous areas, responding to authenticated safety compliance records. PIR sensors sense motion, differentiating between authentic presence and mere nearness. The methodology outlines the necessary hardware and software criteria, procedures for system initialization, evaluation phases, server connectivity setup, access control enactment, and session closure protocols. It details the seamless integration and verification of hardware components, backend connectivity, identity and safety adherence verification, data encoding, and session termination processes. This research aims to upgrade safety surveillance in construction environments, boosting productivity, accuracy, and security. It also underscores the need for further adaptability to various construction settings to advance greater uptake and continuous improvement in workplace safety protocols. 2025 IEEE. -
Reducing Delay and Network Load through Adaptive Threshold-Based Rate Control in IoT Systems
Some of the major challenges in managing IoT networks, which are normally resource-constrained, come from restricted bandwidth, processing power, and energy supplies. Traditional random transmission usually leads to network overload, increased packet delays, and inefficient use of resources. This paper reviews smart rate control mechanisms designed for IoT networks that have limited resources. We analyze and contrast random baseline transmission against threshold-based adaptive control methods by way of extensive simulation runs under realistic network scenarios via the Contiki-Cooja framework. Our experimental results have shown that threshold-based rate control can achieve as much as 31% reduction in average packet delay and 62% reduction in network load when compared with traditional random transmission techniques. Threshold-based rate control represents a deployable and practical solution that properly balances the tradeoff between performance enhancement and computational ease and thus is a good match for real-world IoT deployments over actual resource-constrained networks. Hybrid machine learning, multiobjective optimization, federated learning, and context-aware mechanisms might be potential avenues of future research toward enhancing the performance of IoT systems. 2025 IEEE. -
Evaluating the Impact of Remote Work Benefits on Employee Performance and Retention in IT Companies
The COVID-19 pandemic and other global events that changed the way people work have made remote work more common, especially in IT companies. More and more businesses are using remote employee benefits (REB) to keep good workers, boost productivity, and make people happier at work. Wellness programs, technology allowances, and the option to work flexible hours are all part of this group of benefits. There isn't much real-world proof of how certain REBs affect employee outcomes in IT companies, such as performance, engagement, and turnover rates, even though they are widely used. This study uses a combination of MultiCriteria Decision-Making (MCDM) and other methods to find out how a number of REB affect things. To be more specific, the Analytic Hierarchy Process (AHP) and Fuzzy Logic are used together. The survey had 250 workers from five different multinational information technology companies. The hybrid AHP-Fuzzy model found that the best things for employees to do their jobs better and stay with the company the longest were flexible work hours (0.35), mental health support (0.27), and remote infrastructure allowances (0.21). In comparison, traditional benefits like travel allowances and perks for physical workspaces didn't have as big of an effect. The proposed method was much better than the standard Likert scale and linear regression analysis at giving results that were both accurate and easy to understand. 2025 IEEE. -
Artificial Intelligence Driven Air Quality Prediction for Sustainable Goa
Clean Air is essential for the health and survival of both humans and wildlife. Air pollution has been linked to various serious diseases, including cancer. Rapid industrial growth and increasing population have contributed to rising pollution from transportation, industries, and agriculture. As a result, air pollution has become a major issue, particularly in developing countries like India. To ensure good air quality, accurate and reliable monitoring and prediction are required. Machine Learning (ML) models have shown promise in predicting Air Quality Index (AQI) over traditional methods. This research aims to propose a AQI prediction model using Attention based Bi-directional Long Short-Term Memory (ABiLSTM) to predict AQI in various cities across Goa, India. Data processing methods are used to manage date before providing it into the ABiLSTM model. Daily AQI series from 2022 to 2024 for six cities in Goa- Panaji, Pond, Assanora, Codli, Tilamol, and Tuem are collected and utilized to verify the proposed model. Two models are tested, including BiLSTM and ABiLSTM. Experimental results showed that the ABiLSTM model outperformed BiLSTM model in all cities, reporting lower error values and higher R2 scores. A comprehensive analysis with a set of evaluation indices confirmed that the proposed ABiLSTM model effectively captures the characteristics of the original AQI series and achieves a higher accuracy in AQI prediction. 2025 IEEE. -
Enhanching the Performance Metrics of Overlay Network for QoS in Media Transfer Using Genetic Algorithm
Quality of Service (QoS) of real time video applications is difficult to realize in wireless mobile networks because of the limited resource availability. Software-Defined Networking (SDN) Overlay networks are becoming popular to solve routing, traffic engineering and QoS due to the rapid increase in the adoption and investment in SDN. The SDN market size is projected to grow by a double-digit CAGR within the next decade and reached the low tens of billions USD in 2023, which shows a positive adoption of the industry. Real-time streaming and live content demand have also risen to an all-time high - the live-streaming market is growing at an average rate of about -20-23% CAGR through 2030, and the role of QoS in high-volume media is becoming more and more relevant. 2025 IEEE. -
Deep Learning Analysis of Satellite Images for UN SDG Monitoring in Mauritius' Black River District
This paper proposes an integrated strategy to analyse the progress of selected United Nations Sustainable Development Goals (SDGs 1, 2, and 13) using Earth Observation (EO) data and deep learning (DL) based classification. The research focuses on Mauritius's Black River district, which is facing growing urbanization, agricultural land demand, forest conservation needs, and land degradation. These challenges are closely related to the reduction of poverty (SDG 1) through settlement monitoring, food security (SDG 2) through green farmland analysis, and climate action (SDG 13) through forest cover and bare land tracking. High-resolution satellite images from the Satellogic constellation were pre-processed, classified, and mapped to the SDGs' key land cover categories. A convolutional neural network (CNN) model was trained to distinguish city structures, agriculture farmland, forest land, and barren land, with up to 99% overall precision. DL-based image analysis has the ability to monitor the UN SDGs in specific regions and provide actionable information for the sustainable development plans of small island governments. 2025 IEEE.