Journal of Science Engineering Technology and Management Science

The proliferation of GAN-generated images poses significant challenges to digital media trust. This paper proposes a CNN-based approach for detecting AI-generated images with Explainable AI integration. The model leverages CNN feature extraction to identify subtle artifacts in synthetic images. Grad-CAM and SHAP techniques provide visual and quantitative explanations revealing critical regions the model uses for classification. Experiments demonstrate 95.3% accuracy in distinguishing real from AI-generated images. Grad-CAM visualizations confirm the model focuses on meaningful regions such as unnatural textures and generative artifacts. Despite promising results, limitations include vulnerability to adversarial examples and generalization challenges with novel GAN architectures. The system is deployed as a Django web application enabling real-time image classification with explanations.

Keywords : AI-Generated Image Detection, CNN, GAN, Grad-CAM, SHAP, Explainable AI, Digital Forensics

Author : I. Sireesha, Ch. Sai Charan, M. Hemanth, B. Sandeep

Title : AI-Generated Image Detection with CNN and Interpretation Using Explainable AI

Volume/Issue : 2026;03(03)

Page No : 627-631

Land Use and Land Cover (LULC) classification is critical for environmental monitoring and urban planning. While CNNs achieve high classification accuracy on satellite imagery, their black-box nature limits trustworthiness for policy-making decisions. This paper proposes an interpretable deep learning framework integrating CNN-based classification with SHAP explanations for pixel-level and class-level interpretability. Remote sensing datasets are preprocessed and used to train the model which categorizes land regions into multiple LULC classes. SHAP values explain the contribution of each spectral band to classification outcomes. The system achieves 93.6% overall classification accuracy with SHAP analysis revealing that NIR and SWIR bands are most influential for vegetation and urban class differentiation. The framework is deployed as a Django web application enabling transparent LULC analysis for environmental decision-making.

Keywords : LULC Classification, Remote Sensing, CNN, SHAP, Explainable AI, Satellite Imagery, Django

Author : M. Harish, N. Sowjanya, A. Meghana, S. Dileep Kumar

Title : Interpretable Deep Learning Framework for Land Use and Land Cover Classification in Remote Sensing Using SHAP

Volume/Issue : 2026;03(03)

Page No : 632-636

Machine learning models in financial systems are vulnerable to adversarial attacks and corrupted training data. This paper implements a data enhancement framework for binary classification of relational data to improve model robustness and accuracy. Using the German Credit Dataset, the system trains multiple models (Logistic Regression, Random Forest, Gradient Boosting, SVM, KNN) with preprocessing including feature scaling and categorical encoding. Corrupted data attributes are detected and corrected, and adversarial examples are introduced during training for robustness. The best model is automatically selected and integrated into a Django web application for real-time credit risk prediction. Experimental results show that data enhancement improves average classification accuracy from 71.3% to 79.8% and reduces adversarial vulnerability by 42%, demonstrating effective combination of data enhancement with web-based deployment for robust credit risk assessment.

Keywords :

Author : H. Sruthi, A. Bhavani, D. Girija, K. Harish, K. Srishanthan

Title : Optimizing Relational Databases for High-Performance Binary Classification

Volume/Issue : 2026;03(03)

Page No : 637-641

Most music recommendation systems rely on user history or playlists without considering realtime emotions, reducing personalization effectiveness. This paper presents Emotional-Tune, a vision-driven emotion intelligence system that recommends music based on detected facial emotions. The system captures facial expressions through a camera, processes them using OpenCV for face detection, and classifies emotions (happy, sad, angry, surprise, fear, disgust, neutral) using a CNN model built with TensorFlow/Keras. Detected emotions are mapped to suitable music categories through a recommendation module. The FastAPI backend manages authentication, emotion prediction, and recommendations through secure APIs. Experimental evaluation achieves 87.4% emotion classification accuracy on the FER-2013 dataset, with 91% user satisfaction for music-emotion relevance. The system operates in real-time without additional sensors, demonstrating practical integration of computer vision, deep learning, and recommendation techniques.

Keywords :

Author : POGIRI VENKATANARSIMHULU, POGIRI HARIKA, PAKKI HARIKA, YAGATI BHAVANI, Mrs. B. KUSUMA KUMARI

Title : EMOTIONAL-TUNE: A VISION-DRIVEN EMOTION INTELLIGENCE SYSTEM WITH REAL-TIME MUSIC RECOMMENDATION

Volume/Issue : 2026;03(03)

Page No : 642-648

Traditional election systems using paper ballots or electronic voting machines face challenges including counting errors, tampering, security concerns, and limited accessibility. This paper proposes a blockchainbased online voting system designed to address these shortcomings through decentralized and immutable vote recording. The application leverages blockchain technology to create a transparent, tamper-resistant, and auditable platform for conducting elections. Smart contracts enforce voting rules including one-voteper-voter constraints and voter eligibility verification. The system implements secure voter registration, OTP-based authentication, encrypted vote casting, and blockchain-based vote storage where each vote is cryptographically linked to previous records. Real-time vote counting eliminates manual tallying delays. Evaluation demonstrates 100% vote integrity (zero tampering instances), 99.7% system availability, and 73% reduction in result declaration time compared to traditional systems, while preserving complete voter anonymity.

Keywords :

Author : REYYI HARSHITHA, AMPOLU DURGA,VARANASI SAMATHA,SANAPALA APPALANAIDU,MR.S. BHASKARA RAO

Title : A Digital Voting Architecture Emphasizing Integrity, Trust, and Verifiability

Volume/Issue : 2026;03(03)

Page No : 649-653

Water scarcity has become a critical global challenge due to rapid population growth and the limitations of traditional, delayed monitoring systems. This paper presents ThirstWise Core, a centralized full-stack MERN (MongoDB, Express, React, Node.js) platform designed for real-time water scarcity management through citizen-driven data collection. The system allows users to submit daily localized reports regarding water availability and household consumption. A survival-minimum validation logic ensures reported data aligns with the Government of India standard of 135 Liters Per Capita per Day (LPCD) while enforcing health-safety thresholds. The platform computes a Population-Weighted Stress Index to translate individual household data into macro-level insights. Processed data is visualized via an administrative dashboard using dynamic color-coded indicators (Normal, Warning, and Drought) to identify high-stress regions. Evaluation with simulated data across 15 regions demonstrates 94% stress classification accuracy and 87% faster response time compared to traditional manual survey methods.

Keywords :

Author : SOMARAJU GIRIDHAR, VANKA KALYAN KUMAR, KARANAM DILLESWARI, NETULA MOUNIKA, Mrs.S. ANUSHA

Title : ThirstWise Core: Centralized Platform for Water Scarcity Management

Volume/Issue : 2026;03(03)

Page No : 654-658

Generalized health platforms fail to address individual metabolic differences and fitness goals. This paper presents DeepMetabolix, an AI-driven web-based system providing personalized diet and fitness recommendations. The system analyzes user parameters (age, gender, height, weight, activity level, fitness goals) to calculate BMI and daily caloric requirements using standard metabolic equations (Mifflin-St Jeor). An OpenRouter-based AI chatbot delivers adaptive, context-aware dietary guidance. Graphical visualization and historical tracking improve user engagement. Evaluation with 60 users over 4 weeks demonstrates 92% recommendation relevance, 88% dietary compliance improvement, and 4.4/5 user satisfaction. The system bridges conventional static fitness applications and intelligent personalized health management, offering a scalable solution for metabolic assessment and dietary optimization.

Keywords :

Author : Suravaram Venkatesh, Yerramsetti Madhu Kumar, Muddada Kishor, Palina Lingamurthy, Dr. S. Mouli

Title : DeepMetabolix: Intelligent Architecture for Cross-Sectional Dietary Optimization

Volume/Issue : 2026;03(03)

Page No : 659-664

Prior to the innovation of information communication technologies, social interactions evolved within small cultural boundaries defined by geographic locations. The recent development of communication technologies and social media platforms has considerably transcended the temporal and spatial limitations of traditional communications, creating unprecedented global connectivity. However, the misuse of these social technologies has introduced cyberbullying as a prevalent and harmful form of online aggression that occurs exclusively in digital spaces. Cyberbullying encompasses behaviors including harassment, threats, public humiliation, hate speech, and sustained intimidation through social media posts, comments, and direct messages. Manual content moderation by platform administrators cannot keep pace with the massive volume of social media posts generated daily, creating an urgent need for automated detection systems. This paper presents a comprehensive machine learning-based cyberbullying detection system that analyzes social media text content using advanced NLP feature engineering and multiple classification algorithms. Features extracted include TF-IDF text vectors, profanity count and density ratios, sentiment intensity scores computed using VADER, capitalization ratios, exclamation and question mark frequencies, post length statistics, and user historical behavior patterns. Mutual information-based feature selection identifies the 20 most discriminative features for classification. Random Forest, SVM, and Logistic Regression classifiers are trained and compared on a dataset of 10,000 labeled social media posts with 22% cyberbullying prevalence. Results demonstrate that Random Forest achieves 93.2% accuracy and 0.92 F1-score, with feature selection improving accuracy by 5.7 percentage points. Sentiment intensity and profanity density emerge as the strongest cyberbullying indicators.

Keywords :

Author : CHOPPA SRINU, JADDIDI UMA SRI, AMERAPU MOHAN KRISHNA, BHUSARAPU JYOTHSNA, Mr. SANAPALA BHASKARA RAO

Title : INTELLIGENT APPROACH TO IDENTIFYING CYBERBULLYING BEHAVIOR IN SOCIAL MEDIA USING MACHINE LEARNING

Volume/Issue : 2026;03(03)

Page No : 665-670

Cloud storage has become widely popular due to its flexibility and convenience, but users often have no control over the actual location of their data, which raises concerns about security and trust. To address this issue, the proposed system, LAST-HDFS, integrates Location-Aware Storage Technique with Hadoop Distributed File System. It ensures that sensitive data is stored only within user-defined legal boundaries. The system continuously monitors file migration and replication across cloud nodes to detect suspicious or illegal transfers. File movements are modeled using weighted graph algorithms to group data with similar privacy requirements in the same region. Each cloud node uses a socket monitor to track real-time communication and data transfer activities. Based on this real-time information, the system calculates the probability of illegal file movement. Experimental results in a large-scale cloud environment with 50 nodes and 10,000 migration events show that the proposed system achieves 96.3% detection accuracy with 2.1% false positive rate, demonstrating effectiveness in improving cloud data security and trust

Keywords : Cloud Security, HDFS, Location-Aware Storage, File Migration Detection, Weighted Graph, Data Privacy

Author : BADE KARISHMA, KONCHADA PAVAN, BONI PARVATHI, KALLEPALLI GAYATHRI, Mr. SANAPALA BHASKARA RAO

Title : A LOCATION-AWARE FRAMEWORK FOR DETECTING SUSPICIOUS FILE MIGRATION IN CLOUD

Volume/Issue : 2026;03(03)

Page No : 671-676

Explainable Artificial Intelligence (XAI) has emerged as a critical paradigm in modern financial systems where transparency, accountability, and trust are essential requirements. Traditional artificial intelligence and machine learning models, particularly deep learning techniques, have demonstrated exceptional predictive performance in financial decision-making tasks such as credit scoring, fraud detection, risk assessment, and algorithmic trading. However, these models often function as “black boxes,” providing limited insight into how decisions are derived. This lack of interpretability raises serious concerns among regulators, financial institutions, and end users, especially in high-stakes environments where decisions can significantly impact individuals and organizations. XAI addresses this challenge by incorporating techniques that make AIdriven decisions understandable, interpretable, and justifiable to human stakeholders without compromising performance. In financial decision-making, XAI enables institutions to explain why a loan application was approved or rejected, how risk scores are calculated, or why a transaction is flagged as fraudulent. Such transparency is essential for regulatory compliance with laws such as GDPR, fair lending practices, and ethical AI standards. Moreover, XAI enhances stakeholder trust, reduces bias, and supports better governance by allowing financial analysts and auditors to validate and monitor AI behavior. This study focuses on the role of XAI in enabling transparent financial decision-making, analyzing existing systems, identifying their limitations, and proposing an XAI-driven framework that balances accuracy with interpretability. The research highlights how explainability can transform AI from a purely predictive tool into a responsible decision-support system aligned with ethical, legal, and operational requirements of the financial sector.

Keywords : Software Defined Networks (SDN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Deep Learning (DL), OneDimensional Convolutional Neural Networks (1D-CNN), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Structured Deep Convolutional Neural Network (SDCNN).

Author : BAMMIDI SONIKA, GOKURULA DEEPIKA, BADRI SASIKIRAN, BOMMALI RAJA SEKHAR, B. KUSUMAKUMARI

Title : UNCOVERING MALICIOUS BOTS THROUGH BEHAVIOR-BASED TRANSITION PROBABILITY FEATURES

Volume/Issue : 2026;03(03)

Page No : 677-684

The traditional method of evaluating academic answer sheets, especially descriptive or subjective responses, is a manual and labor-intensive task that often suffers from inconsistencies, evaluator fatigue, human bias, and significant delays in providing results to students. As educational institutions worldwide handle increasing volumes of examinations across multiple subjects and academic levels, the need for an intelligent, efficient, and scalable solution to automate the evaluation process has become critical. This paper proposes an AI-based answer script evaluation system that leverages cutting-edge technologies in Natural Language Processing (NLP), machine learning, and semantic analysis to assess written student responses against pre-defined model answers. The system processes both student and model answers through comprehensive NLP preprocessing including tokenization, stopword removal, and lemmatization before computing TF-IDF vector representations. Semantic similarity between answers is computed using cosine similarity, providing a content-based score that goes beyond simple keyword matching to understand meaning and context. Additionally, the system incorporates spelling and grammar correction modules using language models, and evaluates response coherence and logical structure through sentence flow analysis. The final score is computed as a weighted combination of semantic similarity (60%), grammar quality (20%), and coherence (20%), calibrated against expert human evaluators. Evaluation on 500 answer-key pairs across 5 academic subjects demonstrates a Pearson correlation coefficient of 0.91 with expert evaluators, reducing evaluation time by 92% from an average of 15 minutes per paper to 1.2 seconds while ensuring perfect scoring consistency with inter-rater reliability κ = 1.0.

Keywords : I Answer Evaluation, NLP, Semantic Similarity, TF-IDF, Cosine Similarity, Automated Grading

Author : LANKAVALASA SHANMUKHA RAO, KALLEPALLI RAJYA LAXMI, BOTCHA TEJESWARARAO, HOBBILI KIRAN KUMAR, Mr.S.V.R MURTHY

Title : AI-Driven Intelligent Answer Script Evaluation System

Volume/Issue : 2026;03(03)

Page No : 685-690

Lung cancer kills more people each year than almost any other cancer, and the main reason is late diagnosis. Once a tumour reaches an advanced stage, treatment options shrink dramatically. Screening programs that use Computed Tomography (CT) imaging can catch nodules early, but reading hundreds of CT slices per patient is slow, and two radiologists looking at the same scan often disagree. This paper describes an automated segmentation system built on a deep U-Net architecture that takes raw CT slices as input and returns pixel-level nodule masks without any human in the loop. The encoder half of the network compresses the image into a compact feature representation while capturing broad context; the decoder half reconstructs a full-resolution map; and skip connections carry fine spatial detail directly from each encoder stage to its matching decoder stage so that thin nodule boundaries are not smeared during upsampling. The model was trained on the publicly available LIDC-IDRI dataset using an Adam optimizer with binary cross-entropy loss. Segmentation quality was measured with the Dice Similarity Coefficient and Intersection over Union, both of which reward precise overlap rather than raw pixel accuracy. On the held-out test split the system reached a Dice score of approximately 0.89 and an IoU of approximately 0.82, outperforming classical regiongrowing methods and shallow convolutional baselines. The results suggest this pipeline can meaningfully support radiologists in routine screening workflows.

Keywords : lung cancer detection, nodule segmentation, U-Net, deep learning, CT scan, Dice coefficient, medical image analysis

Author : Balaga. Yaswanth, Mr. P. Satyanarayana, Bodaka. Vijaya,Goona. Dileep, Doddi. Sampath Srinivas

Title : LUNG CANCER SEGMENTATION USING ML & DL

Volume/Issue : 2026;03(03)

Page No : 699-705

Sophisticated malware strains—including ransomware, polymorphic viruses, and advanced persistent threats— continue to outpace conventional signature-driven defenses, demanding a fundamentally more adaptive detection paradigm. This paper introduces a multi-layered malware classification framework that couples dynamic behavioral analysis with an Artificial Neural Network (ANN) classifier and two corroborating detection channels: YARA rulebased pattern matching and crowd-sourced threat intelligence obtained through the VirusTotal API. During controlled sandbox execution, a 28-dimensional feature vector is assembled from file entropy, Win32 API call distributions, portable executable (PE) header attributes, and network-activity indicators. The ANN—a three-layer feed-forward network trained with binary cross-entropy loss and Adam optimization—produces a probabilistic malice score that is subsequently fused with normalized YARA and VirusTotal signals under a weighted risk-scoring formula (ANN 40%, YARA 30%, VirusTotal 30%). Evaluated on a balanced corpus of 12,000 PE executables, the unified system achieves 96.4% detection accuracy, 96.9% precision, and 95.8% recall, surpassing standalone ANN, Random Forest, SVM, and signature baselines by margins of 2.6–18.1 percentage points. End-to-end sample latency averages 4.2 seconds, confirming near-real-time viability. The system is deployed as a Flask web application exposing file-upload, featureentry, and hash-lookup analysis modes, providing analysts with interpretable, actionable verdicts across diverse operational contexts.

Keywords : dynamic malware analysis, artificial neural network, machine learning, YARA rules, VirusTotal API, behavioral detection, cybersecurity

Author : D. Abhisheik, G. Hema Sri Latha, L. Samuel

Title : Dynamic Analysis of Malware Using Artificial Neural Networks with Machine Learning

Volume/Issue : 2026;03(03)

Page No : 706-711

An Artificial Intelligence (AI) transformation is also undergoing a transformation to Software Project Management (SPM), which is equipped with intelligent frameworks that are data-driven and, thus, enhance planning, execution, monitoring, and control across the project lifecycle. The paper presents AI-based strategies, including AI-based Agile frameworks, AI-based DevOps/MLOps integration, smart Decision Support System (DSS) and automatic resource optimization. These models are based on machine learning (ML), predictive analytics, natural language processing (NLP), and real-time monitoring systems to improve the process of the sprint planning, foreseeing risks, cost estimation, and work of the teams. Anomaly detection systems and AI-enabled dashboards assist the proactive decision-making, and MLOps ensure the reliability of the models and ability to be scaled and to meet the governance requirements. The other point that the paper has highlighted is the use of AI in initiation, planning, execution, monitoring, and closure phases which are advantageous in regard to enhanced forecasting accuracy, resource optimization, augmented productivity, and knowledge retention. Even though the opportunities are extremely high, the problems of high-cost implementation, privacy of data, resistance to change and integration challenges have to be addressed to maximize the transformative potential of AI within the framework of contemporary software project environments.

Keywords : Artificial Intelligence (AI), Software Project Management (SPM), Machine Learning, Predictive Analytics and Project Lifecycle Management.

Author : Aravindh Balan

Title : AI-Driven Transformation in Software Project Management: Framework Opportunities, and Challenges

Volume/Issue : 2026;03(03)

Page No : 712-720

Digital learning systems are widely used in modern education, yet most existing platforms provide direct answers without explaining the reasoning process, which reduces transparency and user trust. The objective of this study is to develop a transparent and trust-aware digital learning framework that delivers accurate results along with clear, step-by-step explanations to enhance understanding. The proposed framework follows a multi-layered architecture integrating input processing, feature selection, hybrid reasoning, explanation generation, and trust validation modules. The system is implemented using structured algorithms and evaluated on the Students Performance in Exams dataset. A combination of statistical analysis and rule-based reasoning is applied to generate interpretable outputs, while a validation mechanism ensures reliability through confidence scoring and decision thresholds. The dataset is partitioned into training and testing sets with cross-validation to maintain robustness. Experimental results demonstrate that the proposed model achieves an accuracy of 93.8%, precision of 92.6%, recall of 91.9%, F1-score of 92.2%, and an AUC score of 0.94, outperforming conventional models such as decision trees and neural networks. The system also maintains efficient computational performance with low inference time and stable convergence. The study concludes that integrating transparency, validation, and adaptive feedback significantly improves both learning effectiveness and system reliability. The proposed framework can be effectively applied in intelligent tutoring systems and educational platforms, contributing to more explainable and trustworthy digital learning environments.

Keywords : Transparent Learning Framework, Explainable Knowledge Delivery, Trust-Aware Systems, Educational Data Mining, Student Performance Analysis, Hybrid Reasoning Model, Adaptive Learning, Explainable Artificial Intelligence.

Author : Dr. Anumolu Lasmika1 , Jameer Shaik2 , G. Shirisha3 , Nalukurthi Sumalatha4, Subramanyam Kundili5

Title : A Transparent and Trust-Aware Digital Learning Framework for Explainable Knowledge Deliver

Volume/Issue : 2026;03(03)

Page No : 721-739

India's rich and diverse culinary heritage spans hundreds of regional cuisines, dietary traditions, and cooking styles. Despite significant advances in natural language processing and generative artificial intelligence, no comprehensive system exists that can generate contextually accurate, nutritionally aware, and regionally authentic Indian recipe recommendations tailored to individual user preferences. This paper presents ChefAI India, a generative AI-powered recipe generation system that leverages a fine-tuned GPT-2 medium transformer model augmented with an Indian cuisine knowledge base to produce personalised recipes across six major regional categories: North Indian, South Indian, Bengali, Gujarati, Rajasthani, and other regional cuisines. The model is trained on a curated dataset of 28,400 annotated recipe instances enriched with dietary tags, ingredient substitution mappings, and nutritional metadata. ChefAI India achieves a classification accuracy of 94.3%, precision of 93.1%, recall of 92.8%, and F1-score of 92.9%, outperforming state-of-the-art baselines including Seq2Seq, GPT-2 (vanilla fine-tuned), mBART, and BERT2BERT. User evaluation trials with 120 participants across three Indian cities recorded a satisfaction rating of 4.6/5.0. ChefAI India represents a scalable, culturally grounded solution for personalised Indian cuisine recommendation and recipe generation.

Keywords : Generative AI, Indian cuisine, recipe generation, GPT-2, natural language processing, personalised recommendation, transformer model, transfer learning, dietary customisation, regional food diversity.

Author : Siva Ranjan Das Davalapalli1 , Segari Samuel Rickson2 , Shaik Kharishma3 , Telugu Abilash4

Title : ChefAI India: Personalized Indian Cuisine Recipe Generation Using Generative AI

Volume/Issue : 2026;03(03)

Page No : 740-746

The exponential growth of online review platforms has been paralleled by a surge in fake, incentivised, and spam reviews, fundamentally undermining consumer trust and distorting purchase decisions. This paper presents TrustReview, a hybrid learning framework for automated detection of fake online reviews that synergises a fine-tuned RoBERTa transformer backbone with a multi-modal feature pipeline encompassing linguistic cues, sentiment analysis, reviewer behavioural metadata, and readability metrics. The system is trained on a curated, SMOTE-balanced corpus of 21,540 reviews drawn from Yelp, Amazon, and TripAdvisor, integrated with the ORCA benchmark dataset. TrustReview achieves a classification accuracy of 97.1%, precision of 96.2%, recall of 95.7%, and F1-score of 95.9%, significantly outperforming all evaluated baselines including Naive Bayes (71.4% F1), SVM (77.5%), Random Forest (81.1%), BiLSTM (85.1%), and standalone RoBERTa (89.7%). Deployed as a Flask REST API with sub-second inference latency (0.74 seconds average), TrustReview provides a scalable, real-time solution for review authenticity verification across ecommerce and hospitality platforms

Keywords : Fake review detection, hybrid learning, RoBERTa, sentiment analysis, opinion spam, natural language processing, transformer, SMOTE, reviewer behaviour, e-commerce trust

Author : Subburu Latha¹, S. Vinay2 , G. Chandu3 , K. Ragna4 , Chilupaka Spandhana5 , B. Prathyusha6

Title : TrustReview: Hybrid Learning Framework for Fake Online Review Detection

Volume/Issue : 2026;03(03)

Page No : 747-752