Journal of Science Engineering Technology and Management Science

Consumer Buying Behavior refers to the buying conduct of the closing customer. Many elements, specificities and characteristics impact the individual in what he's and the purchaser in his selection making manner, shopping habits, shopping behavior, the manufacturers he buys or the outlets he goes. A buy selection is the result of every and every one of these factors. An person and a customer is led with the aid of his lifestyle, his lifestyle, his social elegance, his club companies, his own family, his persona, his psychological factors, and so on.. And is influenced with the aid of cultural trends as well as his social and societal environment. By figuring out and understanding the factors that have an effect on their customers, brands have the opportunity to increase a approach, a advertising and marketing message (Unique Value Proposition) and marketing campaigns greater green and more in step with the needs and approaches of contemplating their goal consumers, a actual asset to higher meet the needs of its clients and boom sales.

Keywords :

Author : Dr.R.Hareesh, Chinthala Jyoshna

Title : THE INFLUENCE OF SOCIAL MEDIA INFLUENCERS ON CONSUMER BEHAVIOUR AND PURCHASING DECISIONS AT TATA MOTORS

Volume/Issue : 2026;03(06)

Page No : 623-632

The exponential growth of social media platforms has led to a surge in meme-based communication, with millions of memes shared daily. While many are humorous, a significant portion conveys implicit or explicit hate speech, creating major challenges for content moderation systems. Detecting such harmful content is inherently complex due to the multimodal nature of memes, where meaning arises from the interplay between textual and visual elements. Conventional moderation approaches, including manual review and text-only analysis, are insufficient. Human moderation is time-consuming, subjective, and unable to scale effectively, while text-based models fail to capture visual cues, sarcasm, symbolism, and implicit intent embedded within images. These limitations result in poor classification performance, including high false positives and missed detections of context-dependent hate speech. To address these challenges, this work proposes a multimodal deep learning framework that integrates both visual and textual information for improved hate speech detection in memes. Visual features are extracted using a Vision Transformer (ViT), while textual representations are generated using the eXtreme Language Model (XLNet), enabling robust semantic and contextual understanding. The extracted features are fused into a unified representation and classified using the Sparse Linear Integer Model (SLIM), alongside Logistic Regression, Decision Tree, and K-Nearest Neighbors classifiers for comparative evaluation. Experimental results demonstrate that the proposed approach significantly enhances detection accuracy, reduces misclassification, and improves contextual interpretation. The system supports scalable, real-time deployment and contributes to the development of safer online environments while advancing research in multimodal artificial intelligence.

Keywords : Multimodal Learning, Sparse Linear Integer Model (SLIM), Machine Learning, Vision Transformer (ViT), XLNet, Natural Language Processing.

Author : SK. Mahaboob Basha, T. Neeraj, Madhasu Bhavana, Yallamati Anoop, Ayan Hussain

Title : Deep Semantic and Vision-Transformer Feature Extraction for the Detection of Malicious Patterns in Multimodal Media

Volume/Issue : 2026;03(06)

Page No : 633-642

In today’s digital era, government services must adopt effective automation to enhance accessibility, transparency, and operational efficiency. Conventional certificate issuance systems are largely manual and paper-based, leading to delays, inaccuracies, and potential security vulnerabilities. Citizens frequently experience inconvenience due to repeated visits to multiple government offices, while authorities deal with disjointed processes that are prone to fraud and data mismanagement. To overcome these challenges and modernize certification workflows, a multi-authority e-governance framework has been proposed. This system introduces a centralized, role-based online platform that enables smooth coordination between citizens, healthcare institutions, revenue departments, and administrative bodies. It streamlines the entire lifecycle of certificate handling, including application submission, verification, approval, and secure access, thereby minimizing manual intervention and improving efficiency. Users can register, apply for certificates, monitor application progress, and download approved documents securely using OTP-based authentication. Hospital and revenue officials are responsible for verifying documents within their respective domains, while administrators oversee pending applications, upload approved certificates, and manage notifications for upcoming events. The framework is built using a modern technology stack, incorporating Django for backend development, MySQL for organized data storage, and SMTP-based email services for secure communication. By integrating these components, the system enhances process efficiency, ensures data accuracy, and improves user experience, while maintaining compliance with security and legal standards through authentication and audit mechanisms. Compared to traditional methods, this multi-authority e-governance model offers greater transparency, scalability, and accessibility via a web-based interface. It reduces operational inefficiencies and empowers citizens with quick and secure access to essential documents, ultimately supporting broader digital governance goals focused on trust, accountability, and convenience in public services.

Keywords : E-Governance, Multi-Authority Framework, Certificate Issuance System, MySQL Database, Security Compliance.

Author : P. Anupama, Bandari Akhila, Jaidi Madhumitha, Ganapathi Kowshik Royal, Chandan Reddy

Title : A Smart Digital Governance Framework for Automated Certificate Issuance and Verification

Volume/Issue : 2026;03(06)

Page No : 643-649

The rapid digitalization of healthcare, driven by the adoption of Electronic Health Records (EHRs), connected medical devices, and advanced healthcare software systems, has significantly improved service delivery and operational efficiency. However, this increased reliance on digital infrastructure has also expanded the scope of cyber threats in the healthcare sector. Vulnerabilities present in both modern and legacy systems remain a major source of these threats, making it essential to identify and mitigate risks associated with interconnected medical devices and ICT-based healthcare environments. Threat and vulnerability analysis plays a vital role in reducing potential risks, yet it is challenging due to the large volume of unstructured data, which complicates the identification of meaningful security patterns. To address this challenge, this study presents a Natural Language Processing (NLP)-based framework for automated cyber-threat analysis using machine learning and transformer-based feature extraction. Textual threat reports are preprocessed through tokenization, lemmatization, stop-word elimination, and normalization. Contextual semantic features are then extracted using Decodingenhanced BERT with Disentangled Attention (DeBERTa) embeddings. The generated feature set is evaluated using multiple classifiers, including Greedy Tree Classifier (GTC), Tree-based Algorithm Optimization (TAO) Tree Classifier, K-Nearest Neighbors (KNN), and Gaussian Naive Bayes (GNB), enabling comparative analysis of different learning approaches. The GTC model is selected as the optimal predictor to determine threat category, severity level, and recommended defense mechanisms. The framework is deployed as a web-based system supporting secure data upload, automated prediction, and result visualization, enhancing decision-making and enabling proactive cyber-risk management in healthcare systems.

Keywords : Cybersecurity in Healthcare, Natural Language Processing (NLP), Machine Learning, Transformer Models, DeBERTa, Electronic Health Records (EHR), Greedy Tree Classifier (GTC).

Author : S. Krishna Reddy, Kachu Akshaya, Baki Shivarama Raju, Polampelli Prathyusha, Abdullah Ameen Mohammed Hamid

Title : Smart Healthcare Security: An NLP-Based Framework for Cyber Threat Prediction and Defense

Volume/Issue : 2026;03(06)

Page No : 650-659

The rapid expansion of global connectivity has resulted in the use of over 7,000 languages worldwide, with nearly 60% of internet users engaging in multilingual communication on a daily basis. Despite this growth, recent studies indicate that approximately 40% of multilingual content is either misclassified or insufficiently processed due to the limitations of existing language identification systems. Manual methods for language detection are time-consuming and prone to errors, particularly when dealing with short, informal, or code-mixed text. In addition, traditional computational techniques often lack the semantic capability required to capture subtle linguistic variations present in diverse multilingual datasets. To address these challenges, this study proposes a transformer-based multilingual language identification framework that leverages advanced natural language representations. The process begins with a multilingual dataset that undergoes standard Natural Language Processing (NLP) preprocessing steps, including tokenization, stop-word removal, and lemmatization. Exploratory Data Analysis (EDA) is then performed to examine data distribution and underlying patterns. Semantic feature extraction is carried out using MiniLM, a lightweight transformer-based embedding model designed for high efficiency and accuracy. For comparative evaluation, conventional machine learning models such as Decision Tree Classifier (DTC), K-Nearest Neighbors (KNN), and Naïve Bayes Classifier (NBC) are implemented as baseline methods. The proposed approach utilizes a Random Forest Classifier (RFC), selected for its ability to effectively handle high-dimensional data and leverage ensemble learning for improved predictive performance. This combination enhances multilingual text classification accuracy, enabling reliable identification across short texts, informal expressions, and code-mixed language scenarios. Furthermore, the system is deployed as a Flask-based web application, enabling real-time language detection. This solution demonstrates strong potential for applications in translation systems, multilingual chatbots, and global communication platforms

Keywords : Multilingual Language Identification, MiniLM Embeddings, Natural Language Processing (NLP), Transformer-Based Models.

Author : C. Jyothi Sree, Vani Mallepogu, Ismail Omer Ismail Ishag, Patry Kavya, Esam Hael Saeed

Title : A Translingual Semantic State-Space Modeling Approach for Robust Linguistic Identity Attribution across Multilingual Textual Streams

Volume/Issue : 2026;03(06)

Page No : 660-669

The expansion of digital platforms has resulted in a substantial increase in employee-generated feedback, making the assessment of workforce satisfaction a critical research focus. Conventional methods, such as manual surveys and basic statistical analyses, are often inefficient and inadequate for extracting meaningful insights from complex textual data. Recent developments in Natural Language Processing (NLP) and Machine Learning (ML) have enabled automated approaches; however, many existing solutions face challenges with unstructured text, imbalanced datasets, and multi-label prediction tasks. This study focuses on improving the prediction of key workforce satisfaction dimensions, including work-life balance, skill enhancement, compensation and benefits, job stability, career advancement, and overall job satisfaction, based on employee reviews. Traditional techniques are limited in scalability and often deliver inconsistent results, highlighting the necessity for a more advanced and reliable analytical framework. To address these limitations, the proposed approach combines NLP-based preprocessing, transformer-driven feature extraction using Google PaLM (Pathways Language Model), and the Synthetic Minority Over-sampling Technique (SMOTE). It evaluates multiple machine learning models Quadratic Discriminant Analysis (QDA), Linear Discriminant Analysis (LDA), and Histogram-Based Gradient Boosting (HGB) and compares them with a newly introduced Transformer-Guided Adaptive Model (TGAM). Experimental findings indicate that conventional models achieve accuracy levels between approximately 51% and 56%, whereas the TGAM model attains a perfect accuracy of 100% across all evaluated categories. These results demonstrate the robustness and effectiveness of the proposed system in analyzing complex employee feedback data, supported by comprehensive evaluation metrics and visualization methods for enhanced interpretability.

Keywords : Workforce Satisfaction Analysis, Employee Feedback, Machine Learning (ML), Natural Language Processing (NLP), Transformer-Guided Adaptive Model (TGAM).

Author : K. Vijay Kumar, Chittimalla Akhila, Banavath Shailaja, Polepaka Ramesh, Chiprishetti Nithish

Title : A Transformer-Guided Adaptive Model for Workforce Satisfaction Prediction

Volume/Issue : 2026;03(06)

Page No : 670-682

The global market for conversational Artificial Intelligence (AI) is expected to reach USD 32 billion by 2030, with more than 80% of customer interactions likely to be managed by chatbots. Despite this growth, manual annotation and classification of user intents remain labor-intensive and often inconsistent, limiting the scalability of customer support systems. To overcome these limitations, this study introduces an advanced Natural Language Processing (NLP) framework based on a Customer Support Bitext dataset labeled with multiple intents and categories. The proposed approach begins with NLP preprocessing and Exploratory Data Analysis (EDA) to clean, normalize, tokenize, and analyze data patterns. Subsequently, Miniature Language Model (MiniLM) is utilized to generate lightweight yet semantically rich feature representations. To address class imbalance within the dataset, Synthetic Minority Over-sampling Technique (SMOTE) is applied to create synthetic samples for underrepresented classes. In contrast to traditional models such as Decision Tree Classifier (DTC), KNearest Neighbors (KNN), and Naïve Bayes Classifier (NBC), the proposed system combines Deep Neural Network (DNN)-based feature selection with KNN to improve classification performance. The model is designed to predict two dependent outputs, namely Intent and Category, enabling a deeper contextual understanding of customer queries. Furthermore, the trained model is integrated into a chatbot interface to support real-time intent recognition and automated responses. The proposed framework improves chatbot accuracy, minimizes inconsistencies in annotation, and enhances overall customer experience by effectively handling multi-intent queries. This scalable and efficient solution highlights the effectiveness of integrating advanced embeddings, data balancing techniques, and feature selection methods to advance conversational AI systems in customer support environments

Keywords : Multi-Intent Detection, Customer Support Systems, Miniature Language Model (MiniLM), Natural Language Processing (NLP), Deep Neural Network (DNN).

Author : SK. Mahaboob Basha, Pallapu Ankitha, Kathula Pooja, Osman, B. Vijay

Title : SmartSupport-AI: Multi-Intent and Category Detection Using Lightweight Deep Learning

Volume/Issue : 2026;03(06)

Page No : 683-695

Customer support centers produce a significant amount of textual data through support tickets, which contain valuable information about customer problems, service urgency, and satisfaction levels. Extracting insights from this unstructured data is crucial for enhancing service performance and operational effectiveness. However, conventional methods often depend on manual analysis or basic machine learning approaches such as bag-of-words and Term Frequency–Inverse Document Frequency (TF-IDF). These techniques lack the ability to understand context and semantic relationships, leading to lower accuracy, limited scalability, and difficulty in handling multiple prediction tasks at once, ultimately impacting the quality and speed of customer service. To overcome these challenges, this study proposes a comprehensive analytical framework that combines natural language processing (NLP) with advanced machine learning techniques. The textual data is first cleaned and standardized using common NLP preprocessing steps, including tokenization, removal of stopwords, and lemmatization. Next, transformer-based feature extraction is performed using the eXtreme Language Network (XLNet), which generates contextual embeddings that capture deeper semantic meaning within the text. These embeddings provide a more informative numerical representation, improving the effectiveness of downstream analysis. Various machine learning models, such as Quadratic Discriminant Analysis (QDA), Linear Discriminant Analysis (LDA), Histogram-Based Gradient Boosting (HGB), Stochastic Gradient Descent (SGD), and Nearest Centroid (NC), are then trained on these features to perform multi-target classification. The framework predicts important outcomes, including ticket priority, customer satisfaction scores, and resolution status. In addition, exploratory data analysis is conducted to identify patterns and trends in the dataset. Overall, this integrated approach improves prediction performance, enables scalable processing, and supports more efficient, data-driven decision-making in customer support operations.

Keywords : Customer Support Analytics, Ticket Priority Prediction, Multi-Target Classification, Machine Learning, Natural Language Processing (NLP).

Author : I. V. Prakash, Ale Abhi, Gudtlalwar Chetan, Parthade Kethan, A. Salman, Mynam Vijay Raja

Title : AI-Powered Ticket Insights: Bridging NLP and Ensemble Learning for Smarter Support

Volume/Issue : 2026;03(06)

Page No : 696-707

Attendance management has evolved significantly from traditional manual registers and paper-based record keeping to automated digital systems designed to improve accuracy and efficiency. In educational institutions, offices, and industrial environments, maintaining reliable attendance records is essential for monitoring participation, productivity, and organizational activities. Conventional attendance methods often require considerable manual effort and are prone to issues such as proxy attendance, data entry errors, record loss, and time consumption. Although barcode and password-based systems have been introduced to address some of these challenges, they still suffer from limitations related to user inconvenience, security concerns, and lack of real-time accessibility. These drawbacks highlight the need for a secure, automated, and remotely accessible attendance monitoring solution. To address these challenges, this study presents an IoT-enabled attendance monitoring system that integrates RFID technology, an Arduino-based controller, an ESP8266 Wi-Fi module, a buzzer alert mechanism, an LCD display, and cloud-based data storage. Each authorized user is assigned a unique RFID card that enables quick and contactless identification. Upon successful authentication, attendance information is automatically recorded and transmitted to the Thing Speak cloud platform for real-time storage and monitoring. Invalid card usage triggers an alert through the buzzer, enhancing system security. The proposed approach eliminates manual intervention, improves attendance accuracy, enables remote data access, and ensures reliable record maintenance. The integration of RFID and IoT technologies demonstrates an effective solution for modern attendance management, offering scalability, transparency, operational efficiency, and enhanced monitoring capabilities.

Keywords : RFID, Internet of Things (IoT), Attendance Monitoring System, Arduino Uno, ESP8266 WiFi Module, Thing Speak Cloud Platform, Cloud Data Storage

Author : D Sathish, Mandaloju Sai Balaji, B Dhamodhar, Konakanchi Sampath, Bandapelli Deepika

Title : A Cloud-Connected RFID Architecture for Automated Student Presence Verification

Volume/Issue : 2026;03(06)

Page No : 708-714

The emergence of deep learning-based speech synthesis technologies has enabled the generation of highly realistic artificial audio, commonly known as deep fake audio. These synthetic audio signals can imitate human voices with remarkable accuracy, posing significant threats in areas such as identity theft, misinformation, financial fraud, and digital forensics. Traditional detection methods struggle to identify subtle artifacts present in deep fake audio due to the increasing sophistication of generative models like WaveNet, Tacotron, and GAN-based architectures. This paper proposes an advanced deep learning-based framework for detecting deep fake audio using hybrid neural network architectures. The system integrates Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal sequence modeling. Audio signals are transformed into time-frequency representations such as spectrograms and Mel-Frequency Cepstral Coefficients (MFCCs) to capture discriminative patterns. The model is trained and evaluated on benchmark datasets like ASVspoof, achieving high detection accuracy and robustness across diverse spoofing attacks. Experimental results demonstrate that the proposed approach significantly outperforms traditional machine learning techniques. This research contributes to enhancing audio authentication systems and mitigating risks associated with synthetic media.

Keywords : Deep Fake Audio, Speech Synthesis, CNN, LSTM, MFCC, Audio Forensics, Voice Cloning, ASVspoof

Author : A.Ashok ,Dr G.Vijaya Kumar

Title : DEEP FAKE AUDIO DETECTION USING DEEP LEARNING

Volume/Issue : 2026;03(06)

Page No : 715-723

Human Activity Recognition (HAR) has emerged as a critical component in modern healthcare systems, enabling continuous monitoring and early detection of abnormal behaviors. With the proliferation of wearable devices and IoT-enabled healthcare systems, multimodal data acquisition has become feasible, providing richer contextual information. However, efficiently processing and analyzing such heterogeneous data remains challenging. This paper proposes an intelligent multimodal HAR framework that integrates traditional machine learning models—Decision Tree (DT) and Random Forest (RF)—with a hybrid deep learning architecture, ELM-GRU-AM (Extreme Learning Machine, Gated Recurrent Unit, and Attention Mechanism). The proposed system captures both spatial and temporal features from multimodal signals such as accelerometer, gyroscope, and physiological sensors. The hybrid approach improves classification accuracy, reduces computational complexity, and enhances interpretability. Experimental evaluation demonstrates that the proposed model outperforms baseline methods, achieving superior accuracy and robustness in real-time healthcare monitoring scenarios.

Keywords :

Author : Sk.Azad,Dr G.Vijaya Kumar

Title : INTELLIGENT RECOGNITION OF MULTIMODAL HUMAN ACTIVITIES FOR PERSONAL HEALTHCARE USING RANDOM FOREST, DECISION TREE, AND ELM-GRU-AM ALGORITHMS

Volume/Issue : 2026;03(06)

Page No : 724-732

Ensuring women’s safety in modern society remains a critical challenge due to increasing incidents of harassment, assault, and unsafe environments. Existing safety solutions are largely reactive and rely heavily on manual intervention, which delays response time and reduces effectiveness. This paper proposes an intelligent, integrated women safety system leveraging Machine Learning (ML) algorithms for proactive threat detection and real-time response. The system integrates multi-modal data sources including GPS, accelerometer, gyroscope, and audio signals to monitor user activity continuously. Advanced ML algorithms such as Random Forest, Support Vector Machine (SVM), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) are used to classify safe and unsafe conditions. The proposed system automatically triggers alerts, shares live location, and notifies emergency contacts when abnormal patterns are detected. Experimental evaluation demonstrates high accuracy, low latency, and reliable detection of distress situations. The system significantly enhances women’s safety by providing a proactive, intelligent, and scalable solution

Keywords :

Author : T.Rajesh ,M.Anuradha

Title : EMPOWERING WOMEN: A CREATIVE APPROACH TO INTEGRATED SAFETY WITH MACHINE LEARNING ALGORITHMS

Volume/Issue : 2026;03(06)

Page No : 733-741

Signature verification is a crucial biometric technique widely used for personal authentication in banking, legal documentation, and financial transactions. However, the rise of skilled forgeries poses significant challenges to traditional verification methods. This paper proposes a hybrid signature forgery verification system that integrates Convolutional Neural Networks (CNN) with Histogram of Oriented Gradients (HOG) feature extraction. The extracted features are classified using Decision Tree and Support Vector Machine (SVM) algorithms to enhance accuracy and robustness. The proposed system leverages CNN for deep feature learning and HOG for capturing edge-based structural details, providing a comprehensive representation of signatures. Experimental results demonstrate that the hybrid approach significantly improves classification accuracy compared to conventional methods

Keywords :

Author : A.Sudharsan,Dr P.Chiranjeevi

Title : SIGNATURE FORGERY VERIFICATION USING CNN AND HOG-BASED FEATURES ALONG WITH DECISION TREE AND SVM

Volume/Issue : 2026;03(06)

Page No : 742-748

Demand forecasting is a fundamental component in supply chain management, enabling organizations to predict future product demand and optimize inventory, production, and logistics. Traditional statistical approaches often struggle with nonlinear relationships, seasonality, and high-dimensional data. This paper proposes an advanced forecasting framework integrating machine learning and deep learning models—Random Forest (RF), Gradient Boosting (GB), and Long Short-Term Memory (LSTM). The study emphasizes handling temporal dependencies, feature interactions, and uncertainty in demand patterns. A comparative analysis is conducted using real-world-like datasets, evaluating models based on MAE, RMSE, and MAPE. Experimental findings indicate that LSTM captures temporal dependencies effectively, while ensemble methods ensure robustness and interpretability. The hybrid modeling approach significantly improves forecasting accuracy and scalability in dynamic environments.

Keywords :

Author : K.Teja ,J.Vamsi

Title : DEMAND FORECASTING USING RANDOM FOREST, GRADIENT BOOSTING, AND LSTM ALGORITHMS

Volume/Issue : 2026;03(06)

Page No : 749-756

In today’s data-driven digital economy, understanding customer behavior has become a critical factor for business success. Organizations collect massive volumes of structured and unstructured data from multiple sources such as e-commerce platforms, social media, and transactional systems. However, extracting actionable insights from this data remains a challenging task. This paper presents a comprehensive approach for customer behavior analysis using advanced data mining techniques integrated with Artificial Intelligence (AI)-driven recommendation systems. The proposed system leverages clustering, classification, association rule mining, and deep learning models to identify patterns, segment customers, and predict their preferences. The system enhances personalization, improves customer satisfaction, and increases business profitability. Experimental results indicate that the proposed model significantly outperforms traditional systems in terms of accuracy, precision, recall, and F1-score.

Keywords : Customer Behavior Analysis, Data Mining, Artificial Intelligence, Machine Learning, Recommendation System, Predictive Analytics, Personalization

Author : T.Ashok ,S.Jaya Sri

Title : CUSTOMER BEHAVIOUR ANALYSIS USING DATA MINING TECHNIQUES WITH AI-DRIVEN RECOMMENDATIONS

Volume/Issue : 2026;03(06)

Page No : 757-764

The increasing demand for accessible and efficient healthcare solutions has led to the development of intelligent medical chatbots. Traditional chatbots often rely on static knowledge bases or predefined responses, which limits their ability to provide accurate and context-aware medical advice. This paper proposes a Hybrid Retrieval-Augmented Generation (RAG) chatbot for symptom monitoring that integrates machine learning techniques with large language models (LLMs). The system combines retrieval mechanisms to fetch relevant medical information from trusted datasets with generative models to produce human-like responses. By leveraging both structured medical knowledge and contextual reasoning, the chatbot improves diagnostic assistance and user interaction. The proposed system enhances accuracy, reduces misinformation, and provides real-time healthcare support. Experimental results demonstrate improved performance in terms of response relevance, accuracy, and user satisfaction compared to traditional chatbot systems.

Keywords :

Author : K.JanakiRam ,P.Ramesh Babu

Title : HYBRID RAG CHATBOT FOR SYMPTOM MONITORING USING MACHINE LEARNING AND RETRIEVAL AUGMENTED GENERATION

Volume/Issue : 2026;03(06)

Page No : 765-771

The rapid spread of infectious diseases poses a significant threat to global public health, necessitating advanced predictive systems for early detection and intervention. This paper presents an AI-driven predictive analytics framework that leverages machine learning techniques—Random Forest, Decision Tree, and Artificial Neural Networks—to forecast disease outbreaks. The system integrates historical epidemiological data, environmental parameters, and demographic information to generate accurate predictions. Comparative analysis demonstrates that ensemble and deep learning approaches significantly outperform traditional statistical models in terms of accuracy, precision, and recall. The proposed system aims to assist healthcare authorities in proactive decision-making and resource allocation, thereby reducing the impact of epidemics.

Keywords :

Author : D.Bhargav ,S.Anusha Rani

Title : AI-DRIVEN PREDICTIVE ANALYTICS FOR DISEASE OUTBREAKS USING RANDOM FOREST, DECISION TREE AND NEURAL NETWORK ALGORITHMS

Volume/Issue : 2026;03(06)

Page No : 772-778

Despite the evolution of security measures, SQL injection remains a significant threat to web applications, enabling unauthorized access to databases and valuable information. With the sophistication of attack approaches increasing, traditional signature and rule based detection methods are less effective and thus more intelligent and adaptive detection solutions are required. It is based on a published set of SQL queries, both legit and malicious, from a wide range of web application scenarios from the past few years. The data contains features of the SQL-injection patterns at the keyword level, structural level and character level. Data preprocessing comprises query cleaning, feature extraction, categorical encoding, data normalization, class balancing and duplication removal to enhance the quality and model performance of the data. We create a hybrid detection model to enhance classification performance using various models and evaluate multiple ML classifiers such as DT, RF and XGBoost. The accuracy, precision, recall, f1 score and roc-auc are used to measure performance. The experimental results demonstrate that the hybrid model obtained the best detection effectiveness to be 99.3%, 99.1%, 99.4%, 99.2% and 99.6% for accuracy, precision, recall, F1-score, and ROCAUC, respectively, of the separate classifiers. The proposed framework can enhance SQL injection detection accuracy and provide a good basis for improving the security of modern Web applications.

Keywords : SQL Injection; Machine Learning; Supervised Learning; Query Classification; Web Application Security; Feature Extraction; Decision Tree; Random Forest; XGBoost

Author : Mukkara Sirisha,B. Amarnadh Reddy

Title : A Hybrid Approach to SQL Injection Detection using Machine Learning

Volume/Issue : 2026;03(06)

Page No : 799-807

The proposed work presents an enhanced network intrusion detection framework based on a hybrid Seq2Seq–ConvLSTM architecture integrated with Gated Recurrent Unit (GRU) and Bidirectional learning mechanisms. Modern network environments generate large volumes of sequential traffic data containing complex temporal and spatial dependencies, making accurate intrusion detection a challenging task. To address these challenges, the proposed model extends the conventional Seq2Seq– ConvLSTM framework by incorporating GRU layers to improve computational efficiency and Bidirectional layers to capture contextual information from both past and future network traffic sequences. The GRU component is employed to reduce the computational complexity associated with traditional recurrent neural networks while maintaining effective long-term dependency learning. Compared with conventional LSTM-based architectures, GRU requires fewer parameters, resulting in faster training and inference times without sacrificing detection accuracy. This enhancement enables the model to process large-scale network traffic efficiently and supports real-time intrusion detection requirements. To further improve feature representation, Bidirectional recurrent layers are integrated into the architecture. The Bidirectional mechanism processes network traffic sequences in both forward and backward directions, allowing the model to learn comprehensive temporal relationships and capture sophisticated attack patterns that may not be effectively identified through unidirectional processing. This capability enhances the detection of complex and evolving cyberattacks, including previously unseen intrusion behaviors. The proposed Bidirectional GRU-Integrated Seq2Seq–ConvLSTM framework combines sequence-to-sequence learning for feature encoding, ConvLSTM layers for spatiotemporal pattern extraction, and Bidirectional GRU layers for enhanced temporal dependency modeling. A Flaskbased web application is developed to provide an interactive platform where users can upload network traffic datasets and obtain intrusion classification results in real time. The system supports automated analysis and visualization of detection outcomes, making it suitable for practical cybersecurity monitoring environments. Experimental evaluation demonstrates that the proposed framework outperforms baseline Seq2Seq– ConvLSTM and Random Forest models across multiple performance metrics, including accuracy, precision, recall, F1-score, and detection latency. The results indicate improved robustness against diverse attack categories while maintaining efficient computational performance. These findings suggest that the proposed Bidirectional GRU-Integrated Seq2Seq–ConvLSTM model offers a reliable, scalable, and effective solution for modern network intrusion detection systems operating in dynamic and security-critical environments.

Keywords : Network Intrusion Detection, Hybrid Deep Learning, Seq2Seq, ConvLSTM, Bidirectional Layer, GRU, Temporal-Spatial Features, Real-Time Detection, Flask Deployment, Cybersecurity, Deep Sequential Models, Feature Optimization, Anomaly Detection, Intrusion Prediction, Neural Networks.

Author : B.AMARNATH REDDY 1 ,SURENI.LAKSHMI SUPRIYA 2

Title : Advanced Network Threat Detection Using Sequential and Convolutional Feature Modeling

Volume/Issue : 2026;03(06)

Page No : 808-814

Protecting sensitive information from unauthorized exposure has become a critical challenge in modern cybersecurity environments. Conventional data leakage detection mechanisms often require access to unencrypted content, creating potential privacy and security risks during the inspection process. This paper proposes a privacy-aware data leakage detection framework that integrates Fully Homomorphic Encryption (FHE) with advanced deep learning techniques to enable secure analysis of protected data without revealing its original content. The proposed model, named HE-DLDNet, utilizes a transformer-inspired neural architecture specifically adapted to process encrypted information while maintaining strong confidentiality guarantees. By performing detection directly on encrypted data streams, the framework eliminates the need for decryption during analysis and significantly reduces the risk of sensitive information exposure. Extensive experimental evaluation demonstrates that the proposed approach achieves a detection accuracy of 94.3% while maintaining practical computational performance. Performance comparisons indicate that the framework lowers processing overhead by approximately 73% relative to existing homomorphic encryption-based detection methods. In addition, the system supports near real-time monitoring with response delays below 250 milliseconds, making it suitable for deployment in enterprisescale environments. The results highlight the effectiveness of combining privacy-preserving cryptographic techniques with intelligent learning models to build secure and efficient data leakage detection solutions. This research contributes toward the development of next-generation cybersecurity systems that balance strong privacy protection with accurate threat identification.

Keywords : Privacy-Preserving Security, Fully Homomorphic Encryption, Deep Learning, Data Leakage Detection, Cybersecurity, Transformer Networks, Secure Data Analytics, Network Protection.

Author : Vyshnavi Mattewada,Mrs. J. Shilpa

Title : Privacy-Preserving Data Leakage Detection Using Homomorphic Encryption and Deep Learning

Volume/Issue : 2026;03(06)

Page No : 843-850

The increasing volume of digital financial transactions has made loan fraud detection a significant challenge for modern banking and lending institutions. Fraudulent activities not only result in substantial financial losses but also undermine customer trust and operational efficiency. Although conventional machine learning techniques have been widely adopted for fraud detection, their ability to recognize highly complex and evolving fraud patterns is often limited, particularly when dealing with large-scale, high-dimensional transaction data. To address these challenges, this research presents a Quantum Machine Learning (QML)-based fraud detection framework that utilizes a Quantum Support Vector Machine (QSVM) to improve the identification of suspicious loan transactions. The proposed system combines quantum computing principles with modern web technologies to create an intelligent and scalable fraud analysis platform. By employing quantum feature mapping techniques, transaction data are transformed into a richer feature representation space, enabling the model to capture hidden relationships and intricate behavioral patterns that may be difficult for classical algorithms to identify. The framework integrates quantum circuit execution, classical data preprocessing, and a web-based deployment environment, providing a complete pipeline for real-time fraud assessment and decision support. The model is evaluated using a large collection of loan transaction records containing both legitimate and fraudulent activities. Multiple transaction characteristics, including financial behavior, temporal information, location-based attributes, device-related indicators, and user interaction patterns, are analyzed to generate predictive insights. Experimental findings demonstrate that the proposed QSVM approach achieves superior detection performance compared with several widely used classical machine learning methods. The system delivers high classification accuracy while maintaining a low false alarm rate, making it effective for practical financial applications. Furthermore, the framework supports near real-time transaction analysis with low processing latency, allowing institutions to identify and respond to fraudulent activities promptly. The results highlight the potential of quantum-enhanced machine learning for strengthening financial security and improving fraud prevention strategies. By combining the computational capabilities of quantum algorithms with intelligent predictive modeling, the proposed solution offers a promising direction for next-generation fraud detection systems capable of handling increasingly sophisticated financial threats.

Keywords : Quantum Machine Learning, Quantum Support Vector Machine, Fraud Detection, Financial Security, Loan Transactions, Quantum Computing, Predictive Analytics, Banking Systems, Cybersecurity, Intelligent Decision Support.

Author : Saigommula Ravali,Mrs.N.Sandhya

Title : Quantum Machine Learning for Transactional Fraud Detection in Loan Systems: A QSVM-Based Python Web Application

Volume/Issue : 2026;03(06)

Page No : 851-862

Suicidal ideation detection from social media platforms has become an important research area for identifying individuals who may be at risk and enabling timely intervention. This project focuses on analyzing Twitter posts to detect signs of suicidal thoughts using Natural Language Processing (NLP) techniques implemented in Python with the Natural Language Toolkit (NLTK). The proposed system processes textual data by performing preprocessing tasks such as tokenization, stop-word removal, and feature extraction, followed by sentiment analysis and classification. By examining linguistic patterns, emotional expressions, and contextual cues within tweets, the model categorizes posts into high, medium, or low-risk levels of suicidal ideation. The approach is designed to uncover subtle indicators of psychological distress that may otherwise go unnoticed. This enables healthcare professionals, counselors, and support organizations to take appropriate preventive actions and provide assistance to vulnerable individuals. The project aims to strengthen mental health monitoring, improve early risk identification, and offer a scalable, automated solution that contributes to suicide prevention and the promotion of overall well-being.

Keywords : Suicidal ideation, social media, Twitter posts, natural language processing, NLTK, sentiment analysis, text classification, mental health, intervention, Python, feature extraction, sentiment classification, distress detection, healthcare support.

Author : 1 Paili Rashmitha ,2 J.shilpa

Title : Suicidal Content Detection From Social Media Post

Volume/Issue : 2026;03(06)

Page No : 863-870

The escalating frequency of natural and man-made disasters demands intelligent, automated frameworks capable of detecting emergencies in real time. This paper presents the Unified Emergency Condition Detection (UECD) system, an integrated platform that fuses Internet of Things (IoT) sensor networks, convolutional neural networks (CNN), and deep neural networks (DNN) to detect, classify, and communicate diverse emergency conditions including fires, gas leaks, seismic events, health crises, and vehicular accidents. The proposed architecture employs a hybrid CNN-DNN model trained on structured multi-sensor datasets encompassing temperature, smoke, vibration, heart-rate, and gas-concentration readings, achieving a detection accuracy of 96.7% with a false alarm rate below 1.8%. A Random Forest baseline achieves 93.4% accuracy, while the hybrid deep model surpasses it by leveraging both spatial feature extraction and sequential decision boundaries. Sensor data fusion across heterogeneous modalities substantially reduces ambiguity and false positives. The system generates geo-tagged, severity-ranked alerts distributed via SMS, email, and mobile push notifications to relevant authorities within subsecond latency. An experimental Flask-based web interface demonstrates real-time prediction, dataset management, and model loading capabilities. Experimental results indicate that the UECD system achieves superior performance compared to state-of-the-art singlemodality emergency detection approaches. The system is scalable to smart city deployments, healthcare environments, and industrial facilities, representing a significant advancement toward fully autonomous emergency management infrastructure.

Keywords : Emergency Detection, IoT Sensor Fusion, CNN-DNN Hybrid, Machine Learning, Real-Time Alerting, Smart City, Random Forest, Deep Learning, Multi-Hazard Classification, Flask Web Application

Author : Malkapuram Ramya,Mrs. J. Shilpa

Title : Unified Emergency Condition Detection

Volume/Issue : 2026;03(06)

Page No : 871-876

Despite the evolution of security measures, SQL injection remains a significant threat to web applications, enabling unauthorized access to databases and valuable information. With the sophistication of attack approaches increasing, traditional signature and rule based detection methods are less effective and thus more intelligent and adaptive detection solutions are required. It is based on a published set of SQL queries, both legit and malicious, from a wide range of web application scenarios from the past few years. The data contains features of the SQL-injection patterns at the keyword level, structural level and character level. Data preprocessing comprises query cleaning, feature extraction, categorical encoding, data normalization, class balancing and duplication removal to enhance the quality and model performance of the data. We create a hybrid detection model to enhance classification performance using various models and evaluate multiple ML classifiers such as DT, RF and XGBoost. The accuracy, precision, recall, f1 score and roc-auc are used to measure performance. The experimental results demonstrate that the hybrid model obtained the best detection effectiveness to be 99.3%, 99.1%, 99.4%, 99.2% and 99.6% for accuracy, precision, recall, F1-score, and ROCAUC, respectively, of the separate classifiers. The proposed framework can enhance SQL injection detection accuracy and provide a good basis for improving the security of modern Web applications.

Keywords : SQL Injection; Machine Learning; Supervised Learning; Query Classification; Web Application Security; Feature Extraction; Decision Tree; Random Forest; XGBoost”.

Author : Mukkara Sirisha,B. Amarnadh Reddy

Title : A Hybrid Approach to SQL Injection Detection using Machine Learning

Volume/Issue : 2026;03(06)

Page No : 877-885

Because it provides anonymity, the Dark Web serves as a platform for both legitimate and illicit activity. This work introduces a deep learningbased classification technique that uses TextCNN and Topic Modeling weights to enhance the detection of dangerous Dark Web services. Unlike traditional methods like TF-IDF and Latent Semantic Analysis, which cannot exclude unnecessary data, our approach employs Latent Dirichlet Allocation (LDA) to identify the most crucial topic features. The model is trained using a Kaggle dataset, and preprocessing is carried out to reduce noise and enhance data representation. When LDA-generated topic weights are input into a TextCNN model, classification accuracy is higher than when KNN and Random Forest are employed. An improved CNN2D model with Dropout layers is used to substantially improve classification performance by successfully removing superfluous features. Experimental results show that our approach is highly effective at identifying Dark Web services since it makes use of 90 perfect themes, which significantly increases accuracy.

Keywords : Dark Web Classification, TextCNN, Topic Modeling, Latent Dirichlet Allocation (LDA), Deep Learning, TF-IDF, CNN2D, Dropout Layer, Dark Web Services Detection, Cybersecurity.

Author : J.Kumari 1 , B.Gayathri 2

Title : Enhanced Dark Web Classification Using LDA-Text CNN Feature Fusion and CNN2D

Volume/Issue : 2026;03(06)

Page No : 886-895