Journal of Science Engineering Technology and Management Science

Online social networks represent one of the largest digital ecosystems, with more than 4 billion users worldwide. With this rapid expansion, the risk of maintaining data integrity and security has significantly increased. Social media platforms face numerous security challenges, as users often provide limited personal information, making identity verification difficult. This creates opportunities for scammers to generate fake accounts and misuse them to steal sensitive information, spread misinformation, or harm other users. As a result, detecting fake profiles in social networks has become a complex and important task. This project utilizes various machine learning and deep learning techniques applied to different datasets to identify the most effective model for detection. Additionally, a web-based system is developed to determine whether a given user profile is real or fake. The system is designed with four modules representing major social media platforms that are rapidly gaining users. Each module contains a trained model specifically developed for detecting fake profiles on that platform. Currently, the system has been implemented for Instagram, enabling accurate identification of fake profiles through the developed model with improved performance.

Keywords :

Author : G.Sravanthi, V. VENKATA SAHITHI, V. YASWANTH, P. SINDHUJA, T. AKHILA

Title : Fake Profile Detection on social media Using Python

Volume/Issue : 2026;03(3(1))

Page No : 1-4

With the rapid growth of textual data generated daily from social media, emails, articles, and online platforms, extracting meaningful insights has become increasingly challenging. The Text Analyzer using Machine Learning is a system developed to automatically process, classify, and derive valuable information from large volumes of text data. By applying machine learning techniques, the system performs tasks such as sentiment analysis, topic classification, keyword extraction, and text summarization, enabling users to interpret content more effectively. The system preprocesses textual data using methods like tokenization, stop-word removal, and vectorization to transform unstructured text into structured numerical formats. Machine learning algorithms such as Support Vector Machines (SVM), Random Forest, Naïve Bayes, and Neural Networks are then utilized to analyze and classify the data based on specific objectives. This approach enables automatic detection of patterns, trends, and sentiments within large datasets. By incorporating predictive analytics and intelligent classification, the Text Analyzer supports better decisionmaking, improves content management, and delivers actionable insights across domains such as marketing, education, and social media analysis. The system demonstrates strong accuracy, scalability, and efficiency, making it an effective solution for real-time text analysis and management.

Keywords :

Author : R.Kalpana, M. SAI KOUSHIK, N. SRAVANI, P. HARSHITHA, P. KAVERI

Title : Text Analyzer Using Machine Learning

Volume/Issue : 2026;03(3(1))

Page No : 5-8

In biometric authentication systems, distinguishing between genuine traits and fake or artificially generated samples is a critical challenge, requiring the development of effective security mechanisms. This paper presents a novel software-based fake detection approach that can be applied across multiple biometric systems to identify fraudulent access attempts. The primary goal of the proposed system is to improve the security of biometric recognition frameworks by incorporating liveness detection in a fast, user-friendly, and non-intrusive manner using image quality assessment techniques. The method is designed with low computational complexity, making it suitable for real-time applications. It utilizes 25 general image quality features extracted from a single captured image—the same image used for authentication—to differentiate between genuine and fake samples. Experimental results on publicly available datasets, including fingerprint, iris, and 2D face images, demonstrate that the proposed method performs competitively compared to existing state-of-the-art techniques. The study highlights that analyzing general image quality provides valuable information, enabling efficient and accurate discrimination between authentic biometric traits and spoofed samples.

Keywords :

Author : Easari Parusha Ramu, M. RUTHIKA GOUD, M. DIVYA, V. RAVI KUMAR, D. KEERTHI

Title : Image Quality Assessment for Fake Biometric detection

Volume/Issue : 2026;03(3(1))

Page No : 9-16

This project focuses on developing a software application that enables farmers to sell their agricultural products through a website or mobile platform. The system includes essential features such as product listings, shopping carts, secure payment gateways, and order management functionalities. The primary benefit of an online e-farming platform is its ability to directly connect farmers with consumers. This direct interaction allows farmers to receive fair prices for their produce while enabling consumers to purchase fresh and high-quality products straight from the source. Such a system supports local agriculture and encourages sustainable practices. Additionally, the application can incorporate advanced features like personalized recommendations and order tracking. Personalized recommendations help users discover products based on their preferences, while order tracking provides real-time updates on the status of their purchases, enhancing the overall user experience.

Keywords :

Author : P.Swetha, CH. VAMSHI, D. VENKATESH, K.PAVAN KUMAR, K. SRINIVAS

Title : Harvest Hub: AI Driven Platform for Crop Selection, Price Prediction & Farmer Assistance

Volume/Issue : 2026;03(3(1))

Page No : 17-22

In the modern landscape of rising cyber threats, real-time web traffic monitoring has become essential for identifying and preventing security risks. This study presents the design and evaluation of a Pythonbased real-time web traffic monitoring system implemented on Ubuntu 22.04 LTS. The system combines machine learning-based anomaly detection, real-time traffic analysis, and cryptographic techniques to strengthen cybersecurity monitoring. The system is capable of detecting multiple types of cyber threats, including Distributed Denial of Service (DDoS) attacks with an accuracy of 97.2%, brute-force login attempts with 90.8% accuracy, and unauthorized access attempts with 89.6% accuracy. To improve precision and reduce false positives, optimized anomaly detection methods such as the Isolation Forest algorithm and threshold-based mechanisms are utilized. The system continuously evaluates key network parameters like packet size, request frequency, and response time to identify unusual traffic patterns. An interactive dashboard developed using Flask, along with visualization tools like Plotly and Seaborn, provides real-time insights into traffic behavior, anomaly alerts, and system performance, enabling quick responses to potential threats. Performance testing shows that the system can process up to 10,000 requests per second with an average response time of 150 ms, while maintaining a false positive rate below 10%. Compared to traditional rule-based systems, this approach uses adaptive machine learning models to detect evolving threats more effectively, ensuring improved reliability and efficiency. However, the study also highlights opportunities for future improvements, such as integrating deep learning techniques, deploying the system on cloud platforms for scalability, and incorporating edge computing for faster threat detection. Overall, this work contributes to the advancement of realtime cybersecurity solutions by delivering a high-performance, machine learning-driven monitoring system that enhances security while maintaining operational efficiency.

Keywords : Web Traffic Monitoring, Cybersecurity, Anomaly Detection, Machine Learning, Real-Time Analysis, Ddos Detection, Flask Dashboard, Network Security.

Author : K.Sravani, B. SANDEEP, E. VIKRANTH GOUD, K. JAI ADITHYA REDDY, B. MANIKANTH REDDY

Title : Cloud Security Monitoring

Volume/Issue : 2026;03(3(1))

Page No : 23-32

Waste management has become a significant global concern, as improper disposal methods pose serious risks to both the environment and public health. Conventional waste sorting techniques are often labor-intensive and inefficient, highlighting the need for smarter and more effective solutions. This study presents an AI-based garbage classification system integrated with IoT, combining realtime IoT-enabled sorting with deep learning-driven image classification. The system utilizes a convolutional neural network (CNN) trained on diverse categories of waste to achieve high classification accuracy. Once the waste is identified, an IoT-enabled actuator mechanism automatically segregates it, reducing human effort and improving operational efficiency. To ensure real-time performance and seamless integration with smart city systems, the proposed solution is deployed on edge computing devices. Experimental results demonstrate high accuracy in classification along with efficient waste segregation, indicating the practicality of the system for large-scale implementation. This study highlights the transformative potential of combining AI and IoT in modern waste management, contributing to a more sustainable and circular economy.

Keywords :

Author : N.Manjula, B. THARUN KUMAR, A. SUNNY BHARGAV, G. GANESH KARTHIKEYA, K.VAMSHI

Title : Smart Waste Classification Using Deep Learning

Volume/Issue : 2026;03(3(1))

Page No : 33-42

This paper presents a novel MATLAB/Simulink-based simulation model for a multiport charger designed for light electric vehicles (LEVs) with integrated capability to supply domestic loads. The proposed system employs an Artificial Neural Network (ANN) controller to replace the conventional Proportional–Integral (PI) controller, aiming to enhance system performance under dynamic and nonlinear operating conditions. The multiport converter supports three bidirectional operating modes, namely Vehicle-to-Vehicle (V2V), Vehicle-to-Home (V2H), and Vehicle-to-Grid (V2G), enabling flexible power exchange between electric vehicles, household appliances, and the grid. Unlike the PI controller, which exhibits limitations in handling system nonlinearities and disturbances, the ANN controller is trained to adaptively regulate voltage and current, ensuring optimal performance across all modes. Simulation results demonstrate that the ANN-based control significantly improves transient response, reduces steady-state error, and minimizes Total Harmonic Distortion (THD). Additionally, the controller maintains stable operation and high efficiency during rapid load variations and mode transitions. The comparative analysis confirms that the ANN controller outperforms the conventional PI approach in terms of power quality, robustness, and adaptability. Therefore, the proposed system offers a reliable and intelligent solution for advanced multiport charging applications in modern electric vehicle and smart grid environments.

Keywords : ANN controller, Multiport charger, Electric vehicles, V2G, V2H, Power quality, MATLAB Simulink

Author : Mr. D. JOSEPH KUMAR, Mr. K. SIVA SANKAR, N. L. L. K. PAVAN, CH. S. S. SUBRAHMANYAM, V. RAJESH KUMAR, L. SAI SIMHADRI

Title : AN NOVEL MODEL AND DESIGN OF ANN CONTROLLED MULTIPORT CHARGER SIMULATION

Volume/Issue : 2026;03(3(1))

Page No : 43-55

This project presents a novel energy conversion strategy for standalone photovoltaic (PV) systems employing lithiumion (Li-Ion) batteries, using a Sliding Mode Control (SMC)-based Pulse-Ripple Charging (PRC) technique. The proposed method addresses conventional charging limitations by integrating controlled pulse-ripple current profiles that improve charge acceptance, reduce internal resistance buildup, and enhance battery lifespan. SMC ensures robust dynamic response and system stability under varying solar irradiance and load conditions, making the system resilient to disturbances and modeling uncertainties. The charging scheme is implemented and validated using MATLAB/Simulink, where the system’s performance is analyzed under different irradiance levels and load variations. Simulation results demonstrate a significant improvement in energy conversion efficiency, reduced charging time, and better voltage regulation compared to conventional PI and constant current/constant voltage (CC-CV) methods. The proposed SMC-PRC approach proves to be a viable solution for high-efficiency, reliable standalone PV–Li-Ion energy systems, especially in off-grid and remote applications.

Keywords : Sliding Mode Control, Pulse-Ripple Charging, Lithium-Ion Battery, Standalone Photovoltaic System, Energy Conversion Efficiency, Battery Lifetime Enhancement, MATLAB/Simulink Simulation

Author : Mr. M. SRINU, Mr. K. SIVA SANKAR, P. RAJESH, P. KRISHNA SANKAR, T. M. N. RAGHAVI, B. DURGA PRASAD,S. HEMANTH

Title : AN ENHANCED PULSE-RIPPLE CHARGING STRATEGY USING SLIDING MODE CONTROL FOR STANDALONE PV– LI-ION ENERGY SYSTEMS

Volume/Issue : 2026;03(3(1))

Page No : 56-67

Wireless charging has emerged as a safe and user-friendly alternative to conventional plug-in charging methods for electric vehicles (EVs). Nevertheless, existing wireless charging systems often require additional hardware such as receiver coils and dedicated power converters, leading to increased system complexity, cost, and weight. This paper proposes an integrated wireless charging receiver architecture that leverages existing drivetrain components, specifically the traction inverter and electric motor, to perform receiver-side power conversion. By reusing these components, the need for separate charging circuitry is significantly reduced. The proposed system is designed to limit the circulation of high-frequency currents within sensitive drivetrain elements, thereby minimizing associated losses and enhancing operational reliability. Moreover, the drivetrain inherently regulates the battery charging process, eliminating the requirement for complex communication and control mechanisms from the transmitter side. Simulation studies conducted on a 110 kW EV drivetrain coupled with a 6.6 kW wireless power transfer system demonstrate a maximum efficiency of 94.3% at a coil separation distance of 200 mm. These results validate the effectiveness, efficiency, and practical feasibility of the proposed integrated charging solution.

Keywords : Wireless charging, Electric vehicles, Traction inverter, Dual inverter drives, Wireless power transfer, Charging efficiency, Integrated drivetrain

Author : Mr. I. VIJAY KUMAR, Mr. K. SIVA SANKAR, M. JITENDRA KUMAR, N. G. NAGENDRA REDDY, A. KUMAR, N. KEERTHI KUMAR

Title : DESIGN AND SIMULATION OF AN INTEGRATED WIRELESS CHARGING RECEIVER FOR ELECTRIC VEHICLES WITH DUAL INVERTER DRIVES

Volume/Issue : 2026;03(3(1))

Page No : 68-75

The transition toward a net-zero emission future by 2050 necessitates the large-scale adoption of battery electric vehicles (BEVs) to significantly reduce carbon emissions in the transportation sector. However, one of the primary barriers to widespread BEV deployment is the extended charging time associated with conventional systems. Faster charging demands high-power, high-voltage, and high-current operation, which often leads to efficiency losses across multiple conversion stages. Although traditional AC–DC charging systems provide relatively high efficiency, the cumulative losses from generation to battery charging typically reduce the overall efficiency to below 90%. To address these challenges, DC–DC boost converter technology emerges as a promising alternative for improving energy transfer efficiency. In this study, a superconducting DC–DC double-boost converter (SBC) is proposed and evaluated using MATLAB/Simulink. The system incorporates superconducting inductors based on Bi2223 and MgB? materials. Simulation results reveal that the MgB?-based converter achieves efficiency levels above 95% for output power up to 15 kW. These findings highlight the effectiveness of superconducting technology in enabling ultra-fast, efficient, and high-performance charging solutions for next-generation electric vehicles.

Keywords : Battery Electric Vehicles, Ultra-Fast Charging, DC–DC Converter, Superconducting Inductors, MgB?, Energy Efficiency, MATLAB Simulation

Author : Mr. P. NAGARAJU, Mr. K. SIVA SANKAR, S. RAMA KRISHNA, R.D.A.N. SAI RAM, A. NANI BABU, K. SARAYU, N. MANI KANTA

Title : Artificial Intelligence Based Control of a Superconducting Inductor Assisted DC–DC Converter for Ultra-Fast Electric Vehicle Charging

Volume/Issue : 2026;03(3(1))

Page No : 76-85

Humanemotionrecognitionplaysa vitalroleinenhancinghuman-computer interaction. Thisproject presentsareal-time emotion detection system using deep learning techniques. The system uses a webcam to capture live video input and processes facial expressions to identify emotions such as happy, sad, and angry. The implementation is done using Python with the help of DeepFace for emotion analysis, OpenCV for face detection, and Streamlit for building a userfriendly interface. Themodelutilizes pre-trained deep learningalgorithms to extract facialfeatures andclassify emotions accurately. The system provides real-time output along with emojis and messages, making it interactive and easy to understand. This project demonstrates the effectiveness of artificial intelligence in emotion recognition and its applications in mental health monitoring, smart assistants, and human behavior analysis.

Keywords :

Author : Mrs.K.AshaDevi,S.Lavanya,E.Kiran,B.Dinesh,A.Indu

Title : HUMANEMOTIONRECOGNITIONFROMFACIALEXPRESSIONUSINGDEEPLEARNING

Volume/Issue : 2026;03(3(1))

Page No : 86-94

Sentiment analysis is a Natural Language Processing (NLP) technique used to identify the emotional tone in textual data. This project focuses on analyzing and comparing the performance of various Machine Learning and Deep Learning algorithms for sentiment classification. Models such as K-Nearest Neighbors, Naive Bayes, Decision Tree, Random Forest, Logistic Regression, and Recurrent Neural Network are used to detect sentiments from text data. The dataset is processed using TF-IDF vectorization for machine learning models and tokenization with padding for the deep learning model. The system provides real-time sentiment prediction through a simple Streamlit web interface. It also displays confidence percentages and applies majority voting to determine the final sentiment. This approach helps improve the reliability of sentiment detection by comparing multiple models. The system can be used in applications like social media monitoring, product review analysis, and customer feedback evaluation.

Keywords :

Author : Dr G Yogeswararao, D. Ramya Satya Sree Devi, B. Santhosh Kumar, Ch. Vidya, G. Anand Rao

Title : An Empirical Analysis of Sentiment Detection Using Deep Learning and Machine Learning Techniques

Volume/Issue : 2026;03(3(1))

Page No : 95-101

Face recognition is a key application of computer vision and artificial intelligence used for identifying individuals from images and video streams. This paper presents a real-time face recognition system using deep learning and image processing techniques. The system integrates OpenCV, Haar Cascade classifiers, and Dlib-based face encoding to detect and recognize faces from images, live camera feeds, and videos. A Streamlit-based interface is used for user interaction. The system achieves high accuracy and ensures privacy by performing local processing without cloud dependency. It is suitable for applications such as security systems, attendance monitoring, and access control.The proposed system utilizes OpenCV library for image processing operations, Haar Cascade classifiers for face detection, and the face_recognition library built on Dlib's state-of-the-art face recognition model. The application features a user-friendly web interface developed using Streamlit, enabling users to seamlessly interact with the system through multiple modes: Image Recognition for static images, Live Camera for realtime detection, and Video Recognition for processing pre-recorded videos.The face recognition module employs 128-dimensional face encodings generated using a deep convolutional neural network trained on millions of face images. This approach ensures high accuracy with a tolerance level of 0.5, making the system suitable for various applications including security systems, attendance management, and personal identification. The system demonstrates robust performance across different lighting conditions, facial orientations, and partial occlusions.

Keywords :

Author : Mrs. K.Asha devi, P.Satyanarayana, K.Ramya Anusha, B.Sridhar, R.Sunitha

Title : REAL TIME FACE RECOGNITION FOR VIDEO AND IMAGE PROCESSING USING AI&ML

Volume/Issue : 2026;03(3(1))

Page No : 102-109

Fingerprint recognition is one of the most widely used biometric authentication techniques in modern security systems. However, fingerprint images often suffer from noise, low contrast, smudging, and broken ridge structures, which reduce recognition accuracy. The proposed system focuses on enhancing and reconstructing fingerprint images using image processing and machine learning techniques. The system uses CLAHE for contrast enhancement, Gaussian smoothing for noise reduction, Gabor filtering for ridge enhancement, and morphological operations for improving ridge continuity. Additionally, Support Vector Machine (SVM) is used for classification. The system processes fingerprint images, enhances ridge clarity, and generates improved outputs suitable for authentication.

Keywords : Fingerprint Recognition, Image Processing, CLAHE, Gabor Filter, SVM, Biometric Authentication

Author : N.V.Ashok Kumar,G.Kavya,A.Akhilesh,CH.Ramesh,D.Sathish

Title : FINGERPRINT RECONSTRUCTION AND ENHANCEMENT USING IMAGE PROCESSING AND MACHINE LEARNING APPROACHES

Volume/Issue : 2026;03(3(1))

Page No : 110-117

Cloud infrastructures have become foundational to modern enterprise operations, enabling scalable deployment of services across geographically distributed environments. Despite this widespread adoption, cloud security remains a persistent concern, largely driven by configuration errors that expose sensitive resources to unauthorized access. Conventional Cloud Security Posture Management (CSPM) tools rely on static, rule-driven mechanisms that struggle to adapt to emerging threat patterns in complex multi-cloud deployments. This paper presents a machine learning–based framework for automated detection and classification of cloud misconfigurations. The proposed system employs a Random Forest classifier trained on structured historical configuration records sourced from Amazon Web Services, Microsoft Azure, and Google Cloud Platform. A multi-stage pipeline encompasses data ingestion, categorical feature encoding, temporal feature extraction, model-based prediction, and risk score computation. The trained model is served through a Flask REST API and paired with an interactive web dashboard that provides security analysts with real-time visibility into misconfiguration categories and associated severity levels. Experimental evaluation yields an overall classification accuracy of 94%, a precision of 92%, a recall of 90%, and an F1-score of 91%, confirming the viability of ensemble learning for automated cloud vulnerability management. The modular architecture supports future integration of anomaly detection and Infrastructure-as-Code analysis

Keywords : cloud misconfiguration, cloud security posture management, Random Forest, machine learning, Flask REST API, vulnerability detection

Author : B. Ganesh1, D. Anitha2, V. Gnanasri3, K. Pavan Kumar4, M. Naveen5

Title : AI-Powered Cloud Misconfiguration Detection Using Machine Learning

Volume/Issue : 2026;03(3(1))

Page No : 118-124

Academic institutions face a persistent challenge in delivering timely, accurate, and personalised academic support to large and growing student populations. Conventional Learning Management System (LMS) platforms store considerable volumes of course materials yet lack conversational, context-aware assistance. Existing AI chatbots that rely exclusively on pre-trained Large Language Models (LLMs) frequently produce hallucinated or curriculum-misaligned responses, eroding student trust. This paper presents Student Ease, an Android-based AI powered college helpdesk chatbot that integrates Retrieval-Augmented Generation (RAG) with an LMS backend. The system ingests institutional documents—lecture notes, assignment sheets, and discussion archives—converts them into semantic vector embeddings stored in a FAISS index, and retrieves the most relevant chunks before passing them to an LLM for grounded answer synthesis. A personalization engine tracks each student's query history and learning gaps to provide targeted study recommendations and deadline reminders. System evaluation achieved a Grounded Accuracy of 94.6%, a Citation Precision of 91.8%, a Retrieval Relevance Score of 89.4%, and reduced the hallucination rate from 27% (pure LLM) to 6% (RAG-enhanced). The Flutter/Firebase mobile frontend offers role-based access for students and administrators. Results confirm that combining semantic retrieval with LLM generation substantially outperforms pure LLM baselines in academic helpdesk scenarios.

Keywords : Retrieval-Augmented Generation, Large Language Models, College Helpdesk Chatbot, Android Application, Learning Management System Integration, Vector Embeddings

Author : K. M. Ravi Kumar1, M. Satyanarayana2, T. Avinash3, G. Rama Devi4, N. Swaroop5, V. Balaji Varma6

Title : Android Based AI Powered College Helpdesk Chatbot

Volume/Issue : 2026;03(3(1))

Page No : 125-132

Clickbait represents a deceptive content strategy whereby exaggerated titles and visually manipulative thumbnails are employed to attract user attention without delivering proportionate content quality. This practice gradually diminishes user trust and undermines platform credibility. The present work proposes an automated multimodal clickbait detection framework for YouTube videos that analyzes video titles, thumbnail text extracted via Optical Character Recognition, transcripts, and engagement metrics to classify content as Clickbait or Non-clickbait with associated confidence scores. The system employs Term Frequency-Inverse Document Frequency based text feature extraction combined with supervised learning classifiers including Support Vector Machines and Logistic Regression. Experimental evaluation demonstrates that the proposed approach achieves classification accuracy exceeding 88 percent on standard datasets while maintaining computational efficiency suitable for near real-time deployment. The implementation as a lightweight Flask web application enables practical accessibility for end users seeking to evaluate video credibility before consumption

Keywords : Clickbait detection, YouTube analysis, multimodal classification, thumbnail OCR, machine learning, text mining

Author : D. Kiran Kumar 1, K. Sruthi2, T. Sai3, T. Rohith Kumar4, J. Rajesh5

Title : Clickbait Detection in YouTube Videos Using Thumbnail and Content Analysis

Volume/Issue : 2026;03(3(1))

Page No : 133-139

SQL Injection (SQLi) remains one of the most prevalent and destructive threats to modern web applications, enabling adversaries to bypass authentication, exfiltrate sensitive data, and compromise entire back-end systems. Conventional countermeasures such as Web Application Firewalls (WAFs) and signature-based filters depend on static rule sets that are ineffective against obfuscated, polymorphic, or zero-day payloads. This paper presents an adaptive, data-driven detection framework that leverages Term Frequency–Inverse Document Frequency (TF-IDF) feature extraction coupled with a Random Forest classifier—architected for straightforward migration to Artificial Neural Networks (ANN)—to accurately distinguish malicious from benign SQL queries. The system is deployed as a full-stack web application: a Flask-based REST API exposes a /predict endpoint for real-time classification, SQLite manages user credentials through bcrypt-hashed storage, and a responsive HTML/CSS/JavaScript interface surfaces actionable security alerts. Comprehensive evaluation using accuracy, precision, recall, and F1-score demonstrates that the proposed approach substantially outperforms rule-based baselines while maintaining sub-second inference latency. The modular architecture supports seamless substitution of the classifier with LSTM or CNN models as threat landscapes evolve.

Keywords : SQL Injection Detection, Adaptive Deep Learning, TF-IDF Vectorization, Random Forest Classification, Web Application Security, Flask API

Author : POTLAKAYALA SUDHARSANA RAO , VASUPALLI ABHISHEK2, OMMI NIKHIL KUMAR3, RAYAVARAPU PAVAN KUMAR 4,CHEMBU MANOHAR5

Title : Detection of SQL Injection Attacks Through Adaptive Deep Learning

Volume/Issue : 2026;03(3(1))

Page No : 140-145

Museums are living archives of human civilisation that face a growing challenge— delivering personalised, contextually rich visitor guidance at scale without proportionally expanding human staffing. Existing chatbot solutions built on pattern-matching, intent classification, or knowledge graph paradigms fall short when visitors ask emotional, open-ended, or cross-exhibit questions, and they demand constant manual updates whenever collection content changes. This paper presents a museum chatbot architecture that pairs the generative strength of Large Language Models with the structured, permission-enforced data access of Model Context Protocol servers. The LLM manages natural multi-turn conversation, contextual memory, dynamic tone adaptation, and multilingual generation, while the MCP layer acts as a secure gateway that retrieves verified, real-time exhibit information from the museum database without model retraining or script edits. A four-tier design—visitor interface, LLM processing, MCP communication, and museum database—supports instant content propagation. Evaluation through seven functional test cases confirms stable conversation continuity, reliable multilingual switching, effective restricteddata filtering, and graceful handling of ambiguous queries. The system improves visitor engagement, reduces staff maintenance overhead, and establishes a scalable, privacy-conscious template for AI-powered digital guides across art galleries, cultural centres, and heritage institutions

Keywords : Conversational AI, Large Language Models, Model Context Protocol, Museum Chatbot, Natural Language Processing, Visitor Engagement

Author : G.Rama Devi , N. Amrutha,1P. Sirisha ,1S. Madhu ,1S. Mahesh

Title : Museum Chatbot Using MCP Servers and LLMs

Volume/Issue : 2026;03(3(1))

Page No : 146-152

Road surface deterioration, particularly potholes, poses significant hazards to public safety, increases vehicle maintenance expenditure, and impedes efficient transportation. Traditional inspection strategies relying on manual surveys and complaint-based reporting are reactive, labor-intensive, and unable to provide systematic, real-time coverage of large road networks. This paper presents an automated pothole detection and road damage analysis framework that leverages the YOLOv8 convolutional object detection architecture, integrated with a Flask-based web application, to perform real-time identification and severity classification of potholes from road images, pre-recorded videos, and live webcam streams. The system preprocesses visual inputs using OpenCV, executes inference through a fine-tuned YOLOv8 model, annotates detected regions with bounding boxes and confidence scores, and subsequently classifies each detection into High, Medium, or Low severity based on the proportional area of the pothole relative to the total image frame. Experimental evaluation on the RDD2020 road damage dataset demonstrates progressive reduction in training loss over fifty epochs, with a final mean average precision (mAP@0.5) of approximately 0.57 and a precision of 0.63. The web interface facilitates result visualization, statistical summarization, and JSON-format report export. The proposed system provides a scalable, cost-effective, and accessible solution for intelligent road infrastructure monitoring, with direct applicability to smart city and municipal road maintenance operations.

Keywords : Pothole Detection, YOLOv8, Road Damage Analysis, Deep Learning, Computer Vision, Flask Web Application

Author : B. Ganesh1, K. Jahnavi2, B. Vamsi Naidu3, A. Venkatesh4, P. Venkatesh5

Title : Pothole Detection and Road Damage Analysis Using YOLOv8 Deep Learning

Volume/Issue : 2026;03(3(1))

Page No : 153-156

Gestational Diabetes Mellitus (GDM) is a prevalent pregnancy-related metabolic disorder that poses serious risks to both maternal and neonatal health if undetected at an early stage. Conventional machine learning strategies for GDM detection mandate the aggregation of patient records at a centralized location, introducing substantial privacy vulnerabilities and regulatory challenges. This paper proposes a Federated Learning (FL)-based framework that enables multiple geographically dispersed healthcare institutions to collaboratively train a shared deep learning model without transferring raw patient records outside their premises. Each participating hospital independently trains a local neural network on its private electronic health records, and only encrypted model weight updates are transmitted to a central aggregation server. A Federated Averaging (FedAvg) strategy merges these updates into a continuously refined global model. Clinical features including age, body mass index, blood glucose concentration, blood pressure, pregnancy count, height, and hereditary factors serve as model inputs. Experimental results confirm that the federated approach attains predictive accuracy comparable to centralized learning while providing substantially stronger data privacy guarantees. The system is deployed through a Flask-based REST API and an interactive Streamlit web interface, enabling real-time clinical risk assessment. This architecture offers a scalable and regulation-compliant pathway for privacy-aware healthcare analytics

Keywords : Federated Learning, Gestational Diabetes Mellitus, FedAvg, Privacy-Preserving Machine Learning, Healthcare Analytics, Deep Learning.

Author : P. Jyothsna, S. Sri Varshini, T. Manikanta Swamy, M. Sai, B. Mohit Kumar

Title : Federated Learning for Gestational Diabetes Prediction: A Privacy-Preserving Collaborative Framework

Volume/Issue : 2026;03(3(1))

Page No : 157-161

The Unified Payments Interface (UPI) has transformed digital payment landscapes across India, enabling instantaneous peer-to-peer transactions with over 10 billion monthly transactions by 2023. However, centralized architectures inherent in current UPI systems present vulnerabilities including single-point failures, data manipulation risks, and inadequate fraud detection mechanisms. This research proposes an innovative integration of blockchain technology with UPI infrastructure to establish a decentralized, tamper-resistant transaction validation framework. The implemented system employs cryptographic hashing algorithms (SHA-256), digital signature mechanisms (ECDSA/RSA), and consensus protocols (Proof of Authority) to ensure transaction integrity, authenticity, and non-repudiation. A prototype developed using Python, Web3.js, MongoDB, and Ganache demonstrates enhanced security features including immutable ledger maintenance, real-time fraud detection, and comprehensive audit trails. Experimental results indicate successful transaction validation with minimal latency overhead while providing superior security guarantees compared to conventional centralized systems. The proposed architecture maintains UPI's characteristic transaction speed while introducing blockchain's trust mechanisms, thereby addressing critical security challenges in contemporary digital payment ecosystems.

Keywords : Blockchain, UPI Payments, Cryptographic Security, Digital Signatures, Distributed Ledger, Fraud Detection

Author : G. Jyothi1 , R. Harish Kumar2, N. Surya Baskar3, K. Lokeswari4, A. Jaswanth5

Title : Blockchain for Secure Transactions in UPI Payments

Volume/Issue : 2026;03(3(1))

Page No : 162-167

Advanced cybersecurity issues that modern banking applications face include SQL injection, cross-site scripting, adversarial machine learning attacks, and attempts to compromise credentials. Conventional logging methods undermine incident investigation and regulatory compliance due to mutability issues and a lack of forensic integrity. This study offers a novel architecture that creates immutable evidence collection for cyber defense by fusing blockchain technology with machine learning-based threat detection. The suggested system automatically logs attack evidence on a private blockchain ledger and uses Random Forest classifiers to detect malicious activity in real-time. The implementation makes use of MongoDB for operational data storage, Python based ML microservices, Node.js backend orchestration, and React.js frontend. Using cryptographic verification, smart contracts guarantee the persistence of unchangeable evidence. Experimental validation demonstrates 96% detection accuracy across various attack vectors with sub-second response latency. The blockchain-anchored evidence provides non-repudiable forensic trails suitable for audit and legal proceedings, addressing critical gaps in conventional security information management systems.

Keywords : Adversarial machine learning, blockchain security, cyber defense, immutable logging, smart contracts, banking cybersecurity

Author : B. Ganesh1, V. Karthik2, M.S. Sushma Sailu3, S. Chanukya4, M. Kalyan Kishore5

Title : Immutable Adversarial Attack Evidence Collection for ML Cyber Defense using Blockchain

Volume/Issue : 2026;03(3(1))

Page No : 168-172

This paper presents an intelligent Diet and Fitness Recommender System that integrates Machine Learning (ML), mathematical optimization, and conversational artificial intelligence through the Model Context Protocol (MCP) and Large Language Models (LLMs). The system estimates calories burned during physical activity using a trained Artificial Neural Network (ANN) regression model, taking into account physiological parameters such as age, weight, height, heart rate, body temperature, gender, and workout duration. Predicted caloric expenditure is then used to drive a Linear Programming (LP) optimization module that generates personalized diet plans maximizing protein intake under caloric and macronutrient constraints. An LLM is integrated via MCP as an interactive conversational agent that coordinates backend analytical tools, interprets user queries, and presents recommendations in natural language. The system achieves an ANN Mean Absolute Error (MAE) of 28 kcal, an R² score of 0.93, a Nutrition Constraint Satisfaction Rate (CSR) of 97%, and a System Usability Scale (SUS) score of 82 out of 100 across a 25-participant user study. Experimental outcomes confirm that combining predictive analytics, LP optimization, and conversational AI delivers a more personalized, accurate, and engaging health recommendation experience than conventional approaches

Keywords : Artificial Neural Network, Diet Recommendation, Fitness, Large Language Models, Linear Programming, Model Context Protocol

Author : P. Lavanya 1, M. Chetan Kumar2, S. Sairaj3, M. Bharath4, R. Tharak Rama5

Title : Diet and Fitness Recommender Using MCP and LLM

Volume/Issue : 2026;03(3(1))

Page No : 173-178

The rapid expansion of digital payment ecosystems has intensified the threat of e-commerce fraud, demanding detection systems that are both accurate and interpretable. This paper presents FraudGuard, an intelligent web-based fraud detection application built using the Flask framework. FraudGuard employs a hybrid decision mechanism that integrates rule-based heuristic scoring with machine learning inference from pre-trained ensemble models including XGBoost and a stacking classifier. Transaction risk is assessed across multiple dimensions: transaction amount thresholds, time-of-day windows, geodesic distance between consecutive transaction locations, and international or high-risk location indicators. Raw transaction attributes are transformed into a structured feature vector encompassing log-transformed amounts, temporal flags, and distance-derived indicators. The rule-based engine generates an initial fraud score along with human-readable explanatory reasons, while the machine learning component augments this score with probabilistic estimates from ensemble models. A unified fraud score bounded between 0 and 100 determines a binary fraud classification using a configurable threshold. User authentication is secured via SHA-256 password hashing and session management. Experimental evaluations on representative test scenarios confirm that the hybrid engine consistently assigns elevated scores to suspicious transactions and low scores to legitimate ones. FraudGuard bridges the gap between offline machine learning experimentation and interactive web deployment, serving as both a practical prototype for fraud analysts and an educational reference for integrating ML models into production Flask applications

Keywords : e-commerce fraud detection, hybrid detection system, XGBoost, ensemble learning, geolocation-based risk, Flask web application

Author : M. Meher Tilak¹, P. Gayatri², S. Aslesha³, B. Ganesh Kumar⁴

Title : E-Commerce Fraud Detection

Volume/Issue : 2026;03(3(1))

Page No : 179-184