International Journal of Innovations in Science & Technology
Not a member yet
813 research outputs found
Sort by
Unsupervised Detection of Credential Stuffing and Account Takeover Attempts through User Behavioral Biometrics in Web Applications
Get PDFCredential stuffing and account takeover (ATO) attacks can still be stated as being one of the most ongoing risks of contemporary web applications, as they use reused credentials and defy traditional authentication procedures. Conventional supervised detection approaches rely on labeled attack data (which is frequently non-existent or incomplete) in the real world. In this paper, a credential stuffing and ATO adversarial behavior multi-layer detector is suggested through modeling user behavioral biometrics based on the web-session navigation patterns and chaining information during the login procedure. The framework combines Isolation Forest and Local Outlier Factor to train a normal behavioral distribution and detect deviations that suggest abnormal use of the accounts, which are either automated or reports of hacked accounts. Analysis of three different datasets RBA, MSNBC, and CERT Insider Threat, has shown that the framework is highly detection sensitive with an ROC-AUC of up to 0.936 and an F1-score of 0.842 in login-level anomaly detection, and cross-domain generalization on enterprise data. The findings validate the practicability of unsupervised behavioral modelling as a lightweight defense mechanism that is scalable and resistant to credential abuse. The suggested solution ensures that labelling is reduced, the privacy of users is maintained, and a scalable basic infrastructure is provided to accommodate the adaptive and risk-centric authentication models in massive web applications
Smart Farming with AI: Comparative Evaluation of CNN Models for Tomato Leaf Disease Classification
Get PDFTomato is a major agricultural crop cultivated worldwide; however, its production is severely threatened by a wide range of plant diseases, necessitating accurate and timely detection methods. In recent years, neural network–based computer software and mobile applications have emerged as effective tools for plant disease detection. In this study, three advanced convolutional neural network (CNN) architectures—ResNet-50, DenseNet-121, and InceptionV3—are comparatively analyzed to evaluate their effectiveness in identifying and classifying tomato diseases using the PlantVillage dataset. To enhance model robustness against real-world variability, comprehensive image preprocessing and data augmentation techniques were employed, including rotation, horizontal and vertical flipping, rescaling, shear transformation, and zooming. A systematic hyperparameter tuning strategy was adopted by experimenting with various combinations of learning rates, batch sizes, and optimizers to optimize training performance. Experimental results demonstrate that hyperparameter optimization significantly improves classification accuracy, with the ResNet-50 model achieving the highest accuracy of 98.2%, along with superior F1-score, precision, and recall values. DenseNet-121 and InceptionV3 also exhibited strong performance, although their results were comparatively lower than those obtained with ResNet-50. These findings underscore the effectiveness of transfer learning and fine-tuning strategies in the development of automated systems for plant disease detection and classification. The study highlights the strong potential of CNN-based architectures for scalable and accurate disease detection, offering valuable support to farmers for early diagnosis and improved crop management. Furthermore, the study identifies future research directions, including deployment under real field conditions and the exploration of additional deep learning architectures
Dynamic Behavior of a Magnetized Multi-Component Hybrid Nanofluid on an Oblique Elongating Interface Affected by Extraction and Permeable Media Interactions
Get PDFThe current study explores the mechanism of heat transfer in non-Newtonian Maxwell tri-component nanofluid flow past an inclined stretching sheet embedded in a permeable medium. The electrically conducting nanofluid is considered under the impact of the Lorentz force. The nanoparticles of three types: Silver, Copper, and Ferric oxide, are considered and mixed with the water taken as a base fluid. The proposed phenomenon in the form of differential equations is solved numerically for the numerical outcomes. These results reflect that the Maxwell fluid parameter has an increasing impact on the velocity of the fluid and a decreasing effect on the temperature. The increasing magnetic force effects highlight the increasing trend in temperature of the fluid and the decreasing impact on the velocity of the fluid. The increasing number of nanoparticles has an increasing thermal effect on the fluid. Similarly, the skin friction and rate of heat transfer are dependent functions of pertinent parameters. The differential equations are solved using the exact solver bvp4c
Abstractive Urdu Text Summarization Using Multilingual Transformer Models: A Deep Learning Approach
Get PDFIn contrast to directly copying source text, abstractive text summarization produces short summaries through an understanding of the text. Urdu\u27s low-resource language, which is also characterized by complexities, presents further obstacles. This text investigates the possible extent of deep learning models to automate Urdu text summarization. With respect to the general summary and particular attention to word choice, we try to address the challenges posed by the Urdu language, and we make use of deep learning models for a dataset of Urdu news articles to produce summaries that are accurate and coherent. BERTScore quantitative analysis reveals that the fine-tuned mBART model has an F1 score of 0.497, which is better than mT5 (0.355). As opposed to the most recent Urdu summarization research (2023-2025) in which the majority of reports include ROUGE-based scores, our methodology exhibits a superior semantic consistency and abstractiveness
Preparation of CuSe Thin Films by Chemical Vapor Deposition via Water Splitting for Hydrogen Generation
Get PDFIn recent years, significant research has been done on semiconductor heterostructures to produce hydrogen by water splitting. The absorption of visible light and photoelectrochemical properties of CuO thin film is enhanced by the selenization. The selenization of CuO thin film is done by chemical vapor deposition (CVD) at various temperatures. The structural properties of the prepared samples were carried by XRD and the morphological properties of the prepared film were done by scanning electron microscopy. Optical properties reveal that the bandgap was decreased by increasing the selenization temperature. The solar light to hydrogen conversion efficiency of the CuSe-500oC, CuSe-600oC, and CuSe-650oC films were estimated by using three-electrode cells. It was noticed that CuSe-650oC showed much better STH% compared to pristine CuO thin film
Application of Geospatial Approaches for Evaluation of Urban Growth Pattern and Trend Prediction of Multan City, Pakistan
Get PDFThis research purposes to evaluate the changes in land use, land cover (LULC) in the study area and scrutinize the urban growth trends in Multan City over a period of 30 years, from 1993 to 2023. Moreover, the research utilizes an Artificial Neural Network (ANN) model to implement urban expansion up to the year 2050. To achieve these goals, geospatial systems and approaches are applied. Satellite imagery and remote sensing data from the years 1993, 2003, 2013, and 2023 are analyzed to detect LULC changes. The classification of these images provides valuable insights into the transformation of Multan’s urban landscape over time. A supervised classification technique is primarily utilized to identify specific land cover classes. Landsat 5 data is used for the years 1993 and 2003, Landsat 7 for 2003, Landsat 8 for more recent observations, and Landsat 9 for the latest satellite imagery. The core geospatial model applied in this study is the Cellular Automata–Artificial Neural Network (CA–ANN) model, which is used to simulate and quantify urban expansion. Based on the CA–ANN model results, the urban area in Multan was approximately 154.84 km² in 1993, which expanded to 587.21 km² by 2023. Projections indicate that this urban area will further increase to 992.64 km² by 2030 and could reach 3,184.59 km² by 2050. These findings highlight a significant and rapid urban expansion expected in the coming decades
Vermicompost Effects on Carbon Sequestration of Paulownia Elongata in Agroforestry System
Get PDFIn a controlled agroforestry system, this study evaluated the effects of vermicompost-silt mixtures (30:70, 40:60, and 50:50 ratios) on the growth and carbon sequestration potential of hybrid Paulownia elongata. Compared to other ratios and the control, the 50:50 treatment significantly (p ≤ 0.05) increased plant height (25 inches), stem diameter (0.87 cm), and biomass (13.22 g/plant). According to soil analysis, the 50:50 mixture had the highest carbon stock (6.61 g/plant) and improved potassium (582.67 mg/kg) and organic matter (1.12%) contents. These findings show that Paulownia growth and carbon capture are maximized by balanced vermicompost application, providing a sustainable method for agroforestry in nutrient-deficient soils
Deep Recurrent Neural Network-Based Forecasting of Electricity Consumption and Anomalies Detection
Get PDFThe construction industry is among the greatest fuel consumers of the world and a major source of carbon dioxide. Owing to this, the environmental effect can be minimized when focusing on conserving energy in buildings. Misuse of energy in the effort of equipment and errors of humans in the work is never considered budget. The use of smart buildings will address this problem since it will track the use of energy, detect abnormal behavior, and remind the managers that they are supposed to take energy-conservation actions. The current paper considers the issue of anomaly detection in the hourly electricity consumption level on a real basis and gives a two-step process with a Long Short-Term Memory (LSTM) network. In the first step, there will be forecasting of energy consumption, and, following this, the anomalies will be identified with the assistance of an LSTM Autoencoder. The article draws comparisons between highly complex time-dependent feature extraction algorithms like Rough Autoencoder (RAE), Deep Temporal Dictionary Learning (DTDL). The other algorithms could not perform better than the proposed method, the range of R-squared value was 95.11, MAE was 38.5, the MSE was 2448.94, and the RMSE was 49.49. Besides, the paper evaluates the means through which the AI-based anomaly detection solutions can provide forecasts of the electricity consumption, and the LSTM networks and autoencoders were tested to be more appropriate in forecasting the electricity consumption than the other deep learning algorithms
AI-driven Early Autism Detection and Therapeutic Intervention System
Get PDFEarly identification of Autism Spectrum Disorder (ASD) is crucial for early intervention and improved outcomes. Low literacy and less exposure to computers in Pakistan’s rural areas restrict parents’ capacity to recognize ASD symptoms and receive appropriate interventions. This paper presents an AI-driven, web-based system that fills this gap by providing an accessible autism screening and therapeutic intervention platform. The proposed system integrates machine learning algorithms for symptom-based diagnosis and computer vision for image-based screening. The platform also includes awareness-raising educational content and accessible intervention guidelines for parents. The system is easy to use to ensure accessibility for low-technical-knowledge users. The results indicate that the AI-driven solution enhances the accuracy of diagnosis and provides a scalable solution for early autism screening and awareness in disadvantaged areas
Advances in AI-Based Land Use and Land Cover Classification: A Review of Deep Learning and Remote Sensing Integration
Get PDFThe integration of Artificial Intelligence (AI) with remote sensing has transformed Land Use and Land Cover (LULC) classification, enabling more accurate, efficient, and scalable environmental monitoring. This review synthesizes recent advancements in AI-driven LULC classification, with a focus on deep learning, transfer learning, hybrid approaches, and explainable AI (XAI). Recent studies demonstrate that AI techniques significantly enhance classification accuracy and adaptability across diverse geospatial datasets, supporting applications such as urban expansion monitoring, ecological assessment, reforestation analysis, and real-time land management. Despite these advancements, challenges remain regarding spectral resolution, model interpretability, computational efficiency, and data scarcity. This review highlights these limitations and discusses emerging solutions, including multimodal data fusion, lightweight AI models, and scalable MLOps frameworks. The findings provide insights for researchers, practitioners, and policymakers to guide future work in sustainable land management and environmental monitoring