Sentiment analysis and sarcasm detection as become an important area in natural language processing (NLP) due to growth of e-commerce and social media platforms. Customers give feedback through reviews which helps to understand the contextual meaning and sentiment present in the text.The system integrates DistilBERT and CNN-LSTM hybrid model where CNN (Convolutional Neural Network) used to extract the local features, and LSTM (Long Short-Term Memory) learns sequential and long-term dependencies in the text, and DistilBERT is used to capture contextual embeddings in the text. The proposed system processes the reviews through preprocessing techniques including text cleaning, negation handling, and tokenization. Experimental results shows that the hybrid model achieves high accuracy compared to standalone and architecture architectures. The proposed model is applied in e-commerce platforms, recommendation systems and customer feedback analysis for better decision making.
Malware has become increasingly sophisticated, making traditional signature-based detection insufficient for identifying new and obfuscated threats. This paper presents an ensemble learning-based malware detection system using Bagging (Random Forest), Boosting (AdaBoost and Gradient Boosting), and Stacking techniques. The dataset contains 100,000 Android application samples equally divided into malware and benign classes. Data preprocessing includes label encoding, feature scaling, and removal of irrelevant attributes. Experimental results demonstrate that ensemble learning improves accuracy, precision, recall, and F1-score, providing a robust solution for modern cybersecurity applications.
This paper introduces a better method of predicting heart disease, combining advanced preprocessing methods of data with a powerful ensemble learning model. This dataset underwent a thorough cleaning process, and the issue of class imbalance was solved with the help of the Synthetic Minority Oversampling Technique (SMOTE), which was applied solely to the training dataset to ensure that no information is leaked when the model is being tested. The ANOVA F-test was used to select the most significant clinically features with SelectKBest method, which identified eight clinically significant features, such as age, sex, chest pain type, maximum heart rate, exercise-induced angina, ST depression, ST slope, and number of major vessels. In order to improve predictive performance, a stacking ensemble model was constructed where the base learners were Random Forest and Extra Trees and Logistic Regression was the meta-learner to merge the results of the base learners. Moreover, the classification threshold was lowered to 0.6, to be more accurate to the medical diagnosis, which is a costsensitive issue. The proposed model had a high accuracy of 97.88, high precision, recall, and F1-score, and a ROC-AUC of 99.76, which is higher than the traditional methods. SHAP (SHapley Additive exPlanations) was used to provide model transparency with the KernelExplainer method and the high-dimensional results were rolled into a two-dimensional form to be visualized effectively. The most significant predictors identified in the analysis are sex, exercise-induced angina, age, number of vessels, and ST slope, which is in line with the known clinical knowledge. In general, the findings indicate that systematic preprocessing, ensemble learning, and explainable AI can be used to achieve a high level of prediction reliability and be practically applicable in cardiovascular healthcare. Future efforts will be directed at additional optimization of the pipeline and integration of realtime clinical data to maintain and improve predictive performance.
The rapid proliferation of Internet of Medical Things devices, including wearable sensors, implantable monitors, and cloud-interfaced diagnostic systems, has introduced a critical and expanding attack surface alongside its transformative clinical benefits. IoMT environments are characterized by resource-constrained hardware, heterogeneous protocols, and highly sensitive patient data transmission, making them prime targets for botnet-driven distributed denialof-service attacks, data exfiltration, and network intrusion. Existing security frameworks treat anomaly detection and audit logging as loosely coupled modules, lacking the latency guarantees, privacy protections, and regulatory alignment required for live hospital infrastructure. This paper proposes a tightly integrated, privacy-preserving security framework combining a Transformer-based Autoencoder for real-time anomaly detection with a Hyperledger Sawtooth permissioned blockchain for immutable audit logging. The Transformer Autoencoder is trained exclusively on benign traffic from the N-BaIoT dataset using an unsupervised reconstruction-loss paradigm, learning a compact latent representation of normal behavior that enables detection of deviations induced by sophisticated attack traffic. Upon detection, a middleware pipeline constructs a privacy-preserving metadata payload comprising a device identifier, timestamp, anomaly score, and SHA-256 cryptographic hash of the raw input, committing it to the distributed ledger without exposing any sensitive data on-chain. Empirical evaluation demonstrates 98.6% accuracy, 97.1% precision, and 98.4% recall on the N-BaIoT test set, outperforming eight contemporary baselines whilst maintaining 19ms inference latency. The blockchain layer sustains commit latency below 90 milliseconds at a two-kilobyte payload, with linear throughput scalability from 20 to 48 transactions per second. The off-chain hash-only storage model resolves the blockchain-GDPR paradox, satisfying both Article 17 right-to-erasure obligations and HIPAA transmission security standards through deliberate architectural design rather than post-hoc compliance measures.
As quantum computing technology transitions from theoretical frameworks to scalable physical implementations, the "Harvest Now, Decrypt Later" (HNDL) strategy has emerged as a primary existential threat to long-term digital security. Adversaries are currently intercepting and archiving encrypted sensitive data with the intent of performing retroactive cryptanalysis once cryptographically relevant quantum computers (CRQC) become available. This paper addresses the vulnerability of current Public Key Infrastructure (PKI), specifically RSA and Elliptic Curve Cryptography (ECC), which are susceptible to Shor’s algorithm. We propose "Aethelgard," a high-performance hybrid cryptographic framework that nests NIST-standardized Post-Quantum Cryptography (PQC), specifically ML-KEM (Kyber), within established classical key exchange protocols. By utilizing a dual-layered Key Encapsulation Mechanism (KEM), our framework ensures that data remains confidential even if one of the underlying mathematical primitives is compromised. We evaluate the framework’s efficacy through rigorous network simulations, measuring computational overhead, handshake latency, and packet fragmentation across diverse traffic profiles. Experimental results indicate that while the hybrid approach increases the public key size to approximately 1.2 KB, the overall handshake latency increases by a manageable 18% compared to classical-only protocols. This research demonstrates that hybridizing cryptographic primitives is the most viable path toward achieving immediate "crypto-agility" and mitigating the silent threat of HNDL without sacrificing the reliability of current production environments.
Tuberculosis (TB) is one of the most dangerous infectious diseases in the world and early detection is very important for saving human lives. Manual reading of chest X-rays by radiologists is time consuming and not always available in remote areas. This project is made to reduce this problem by building an automatic system for detecting TB from chest X-ray images using deep learning. The system uses a convolutional neural network model to classify X-ray images as TB-positive or TB-negative. Pre-trained models like VGG16 and ResNet50 are used with transfer learning to improve accuracy on limited medical data. Image preprocessing techniques like histogram equalization, resizing, and data augmentation are applied before training. Flask is used for building the web interface and MongoDB is used for storing patient records and prediction results. The system can automatically analyze uploaded X-ray images and give prediction result with confidence score. After testing, the system shows high accuracy and can help doctors and health workers in early diagnosis of tuberculosis.
Network Intrusion Detection Systems are very good at detecting malicious network traffic; however, they do a poor job of providing analysts with the information to diagnose a breach or make decisions about what actions should be taken. This research describes an AI-based autonomous breach diagnostic system. The proposed autonomous breach diagnostic system uses machine learning algorithms, XGBoost and HistGradientBoosting, in combination with data from multiple sources, uncertainty-based classification, SHAP explanations, incident clustering/grouping, and corroborated evidence to determine which alerts may represent malicious activity. The proposed autonomous breach diagnostic system has been tested using three different datasets: CICIDS-2018, CICIDS-2017, and UNSW-NB15. On CICIDS-2018, the proposed model achieved F1 = 0.9892 and AUC-ROC = 0.9990. In addition, it reduced 405,032 raw detections down to 8,702 diagnosed breaches that included 499 high-priority alerts.
Oral cancer remains a major global health burden, accounting for significant morbidity and mortality worldwide. Despite considerable advances in treatment strategies, survival rates remain relatively low due to delayed diagnosis and limited access to early detection facilities. Conventional diagnostic approaches, including visual examination and biopsy, continue to serve as the standard methods for diagnosis. These traditional invasive methods are time taking and painful and often identify lesions only after substantial disease progression. Consequently, there is a growing need for innovative, non-invasive, accurate, and cost-effective diagnostic technologies capable of identifying oral cancer at its earliest stages. Recent advancements in biomedical engineering, molecular biology, artificial intelligence (AI), and imaging technologies have significantly transformed the landscape of oral cancer detection. Advanced optical imaging techniques, including fluorescence imaging, narrow-band imaging, and con-focal microscopy, have enhanced the visualization of precancerous and malignant lesions. Simultaneously, molecular diagnostic approaches utilizing salivary biomarkers, circulating tumor DNA (ctDNA), proteomics, and metabolomics have emerged as promising tools for non-invasive cancer screening. Artificial intelligence-based diagnostic systems, portable point-of-care devices, and telemedicine platforms further support early diagnosis and improved accessibility to healthcare services. This review discusses recent advancements in oral cancer detection technologies, their clinical applications, advantages, limitations, and future prospects.
ABSTRACT: Cloud storage platforms like Google Drive are where people store a lot of evidence these days. Google Drive has over three billion users. This makes it a key place to look for evidence in investigations. The problem is that the tools we use to look at evidence like FTK, Autopsy and EnCase were made for looking at evidence on storage media. They do not work well with storage. They also do not have open-source support for getting evidence from the cloud. This is why we made Cloud Track. Cloud Track is an open-source tool that uses the Google Drive API to look at all the files on Google Drive. It checks the files to make sure they have not been changed. It does this by using something called SHA-256. It also keeps a record of what has been done to the files. This record is like a chain of custody. Cloud Track can also look for files that're not normal. It does this by using something called an Isolation Forest. This helps us look at a lot of files quickly. We can also export the evidence in a way that's safe and secure. We use something called AES-256-GCM to encrypt the evidence. We have an interface that makes it easy to use Cloud Track. The interface has eight tabs that help us investigate. We have tested Cloud Track with, up to 1,000 files. It works well. It can find the files we are looking for 94% of the time. It never finds something that's not there. Cloud Track follows all the rules for looking at evidence. It follows the NIST SP 800-86 model. It also follows the volatilityordering requirements of RFC 3227. This means that the evidence we get from Cloud Track can be used in court. Cloud storage platforms like Google Drive are where people store a lot of evidence these days. Google Drive has over three billion users. This makes it a key place to look for evidence in investigations. The problem is that the tools we use to look at evidence like FTK, Autopsy and EnCase were made for looking at evidence on storage media. They do not work well with storage. They also do not have open-source support for getting evidence from the cloud. This is why we made Cloud Track. Cloud Track is an open-source tool that uses the Google Drive API to look at all the files on Google Drive. It checks the files to make sure they have not been changed. It does this by using something called SHA-256. It also keeps a record of what has been done to the files. This record is like a chain of custody. Cloud Track can also look for files that're not normal. It does this by using something called an Isolation Forest. This helps us look at a lot of files quickly. We can also export the evidence in a way that's safe and secure. We use something called AES-256-GCM to encrypt the evidence. We have an interface that makes it easy to use Cloud Track. The interface has eight tabs that help us investigate. We have tested Cloud Track with, up to 1,000 files. It works well. It can find the files we are looking for 94% of the time. It never finds something that's not there. Cloud Track follows all the rules for looking at evidence. It follows the NIST SP 800-86 model. It also follows the volatility-ordering requirements of RFC 3227. This means that the evidence we get from Cloud Track can be used in court.
Personalized nutrition planning is essential for maintaining healthy lifestyles, managing weight, and improving overall well-being. However, generic diet recommendations often fail to consider individual nutritional needs, dietary preferences, regional food habits, and practical implementation. To address these challenges, NutriPlan is proposed as an intelligent personalized Indian diet and grocery planning system. The application utilizes a structured dataset of Indian meals containing nutritional values, dietary tags, regional categories, and ingredient information to generate customized recommendations. Using a rule-based and AI-inspired framework, the system combines profilebased filtering, fuzzy nutritional scoring, and TOPSIS-based multi-criteria ranking to recommend suitable meals. Based on user goals and preferences, NutriPlan creates a seven-day meal schedule covering breakfast, lunch, snacks, and dinner. The system also generates consolidated grocery lists, tracks progress, and evaluates plan effectiveness through macro-nutrient summaries, consistency scores, weekly trends, and achievement indicators. By providing transparent, scalable, and explainable recommendations, NutriPlan offers a practical solution for real-world personalized nutrition and diet management.
Literature acts as a compelling medium for reflecting societal structures, norms, and changes, especially concerning gender interactions. The gender perspective examines how an individual's gender influences their relationships, social standing, and available opportunities. Gender impacts every aspect of an individual's or society's private, social, and economic life, as well as the various roles that society assigns to both men and women. This research provides a critical analysis of how Indian literature depicts gender dynamics, focusing on the changing representation of masculinity, femininity, and non-binary identities throughout different eras. By examining the works of prominent Indian authors such as Sudha Murty, Ruskin Bond, Mahasweta Devi, and Amitav Ghosh, this study explores the intersectionality of gender, culture, and power within literary narratives. It underscores how traditional texts either uphold or contest patriarchal values, whereas contemporary literature presents more nuanced insights into gender roles and identities. Characters defined as male and female are often depicted in contrasting ways within literary works. The research also examines the historical progression of gender themes in Indian literature, tracing transitions from mythological and classical representations to modern discussions of gender fluidity and resistance. This study aims to unveil the profound consciousness embedded in literary works through a comparative approach, revealing how gender interactions both influence and are influenced by Indian society.
Liver disease continues to be a major global health concern due to its silent progression and delayed diagnosis. This study proposes an efficient and interpretable machine learning framework for early liver disease prediction by integrating advanced preprocessing, Dual Statistical Feature Optimization (DSFO), and ensemble learning. Missing values are handled using iterative imputation techniques, while class imbalance is addressed using SMOTE to ensure balanced data distribution. The proposed DSFO method combines Chi-Square dependency analysis and ANOVAbased variance evaluation to identify the most significant clinical features. Unlike conventional stacking approaches, the model employs a simplified ensemble strategy using LightGBM and Extra Trees Classifier combined through soft voting, improving generalization while reducing model complexity. The framework achieved an accuracy of 97.96% and ROCAUC of 0.9894, demonstrating strong predictive capability. To enhance interpretability, SHAP is used to provide insights into feature contributions. Results reveal that biochemical attributes such as bilirubin and liver enzyme levels play a dominant role in prediction. The proposed system offers a reliable and transparent solution suitable for real-world clinical applications.
Reconstructing visual experiences from non-invasive brain signals is a fundamental goal in Brain Computer Interface (BCI) technology. This project aims to create a system that decodes raw Electroencephalography (EEG) signals and reconstructs a photorealistic approximation of the image a person is viewing. This "visual mindreading" moves beyond simple classification to the direct generation of pixel-level content. This task is exceptionally challenging due to the high noise and low spatial resolution inherent in EEG data. Current brain decoding research has had more success with fMRI, which offers high spatial resolution, but it is non-portable and expensive. EEG-based reconstruction is far more practical but also more difficult, representing an "extreme modality gap." Early attempts, often using Generative Adversarial Networks (GANs), have struggled to bridge this gap, producing low-resolution, blurry, or non-representative images that fail to capture the rich semantic detail of the original visual stimuli. This project introduces Cognitive Canvas, a novel dual-stage framework. First, a Transformer based EEG Encoder, trained with contrastive learning, will be designed to extract robust semantic feature vectors from the noisy, multi-channel time-series data. Second, a state-of-the-art Latent Diffusion Model (LDM) will be conditioned on these EEG feature vectors instead of text prompts to generate the final photorealistic image. This disentangled approach is designed to effectively translate abstract brain activity into a coherent visual space. The core implementation will use Python, PyTorch, and the Hugging Face diffusers library. The EEG Encoder will be a custom Transformer. Signal processing and feature extraction. Model alignment will enhance CLIP style contrastive learning. The system will be trained and evaluated on public EEG-visual datasets, primarily Mind Bigdata.
ProTask AI: GenAI Integrated Task Manager is an intelligent task management system that combines generative AI, natural language processing, and artificial intelligence (AI) technologies to increase productivity. By enabling users to create, amend, and manage tasks using natural conversational inputs rather than tedious form-based interactions, the system gets beyond the drawbacks of conventional task managers. Through the Ollama framework, it makes use of a locally deployed Large Language Model (LLM), guaranteeing data privacy, lower latency, and real-time processing without depending on outside cloud services. The main feature is a conversational interface that allows users to give unstructured commands using both text and speech-to-text input. After interpreting user intent, the AI pulls task details—such as description, deadline, priority, and reminders—and transforms them into structured JSON format so that they may be processed. Secure REST APIs created with FastAPI manage this data and enable effective CRUD operations. Accuracy and consistency are improved through the integration of LangChain with rapid engineering, context management, and structured data extraction. FastAPI is used in the backend for scalability and high speed, and MongoDB is used to safely store task and user data. Along with visualization options like calendar views and progress charts for efficient tracking, the system also provides automated email and browser notifications for timely reminders. In contrast to current systems, ProTask AI ensures modularity, maintainability, and adaptability with a tiered architecture that includes User Interface, Application, AI, Database, and Notification levels. The system offers a userfriendly, effective, and privacy-focused solution that streamlines job management and boosts overall productivity by fusing conversational AI, speech recognition, and intelligent automation.
People with visual impairments often face difficulties while navigating unfamiliar indoor and outdoor environments due to obstacles, changing surroundings, and limited awareness of nearby hazards. To address these challenges, this paper presents a Deep Learning Driven Augmented Reality (AR) Navigation System with Audio-Haptic Feedback that provides real-time navigation assistance and supports independent mobility. The proposed system uses YOLOv8 (You Only Look Once Version 8) for object detection and MiDaS (Mixed Datasets for Monocular Depth Estimation) for depth estimation to identify pedestrians, vehicles, furniture, obstacles, and other objects in the user's surroundings. Live video frames captured through a mobile device camera are continuously processed using a FastAPI backend to analyse the environment and determine safe movement directions based on obstacle location and distance. To improve accessibility, the system provides voice guidance in English, Hindi, and Telugu, along with vibration-based haptic alerts that help users respond to nearby obstacles and potentially unsafe situations. AR overlays such as bounding boxes and directional indicators are displayed on the screen to highlight detected objects and navigation regions, making the system useful for partially sighted users and system monitoring. Temporal smoothing techniques are applied to improve the consistency of navigation guidance and reduce unstable decisions during movement. Experimental evaluation demonstrated reliable obstacle detection and effective navigation decision generation under different environmental conditions. By combining deep learning, computer vision, speech interaction, and multimodal feedback, the proposed system offers a practical, affordable, and user-friendly assistive solution that helps visually impaired users navigate more safely, confidently, and independently.
Undetected plant diseases remain a persistent threat to agricultural productivity worldwide. This paper presents Leaf Guard-AI, a browser-accessible diagnostic tool built on convolutional neural network (CNN) technology to classify leaf diseases from uploaded images. Implemented using Tensor Flow/PyTorch and deployed via a Gradio web interface, the system requires no software installation. Beyond disease labelling, it provides confidence estimates and crop-specific remediation guidance. Evaluation on a held-out test partition confirmed classification accuracy exceeding 90% with sub-two-second inference latency, validating the tool for on-farm deployment.
Internet of Medical Things (IoMT) applications require collaborative learning across healthcare institutions while ensuring patient data privacy. Traditional centralized learning approaches require sharing sensitive medical records, increasing privacy and security risks. Federated Learning (FL) enables distributed model training by exchanging model parameters instead of raw data, but its performance is affected by client dropouts and communication failures. This paper proposes a Resilience-Enhanced Federated Learning Framework for IoMT applications that improves the reliability of collaborative learning under unstable network conditions. The framework incorporates Federated Averaging (FedAvg), resilient aggregation using historical model updates with staleness decay, and quantized model updates to reduce communication overhead. The proposed model was evaluated using a heart disease dataset distributed across multiple healthcare clients. Experimental results demonstrate that the framework achieves 95.72% accuracy while maintaining stable model convergence during client failures. The proposed approach provides a secure, privacypreserving, and fault-tolerant solution for distributed healthcare applications.
The evolution of malware led to the emergence of polymorphic variants capable of continuously modifying their code structure to evade traditional detection mechanisms. Such adaptability renders conventional signature-based and heuristic approaches largely ineffective, creating a need for intelligent and adaptive security solutions. This research presents a framework for real-time detection of polymorphic malware, integrating deep learning and synthetic data augmentation to enhance robustness and generalization. The proposed system leverages Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Attention LSTM along with other algorithms like Random Forest, Voting Classifier, Stacking Classifier and Naive Bayes to jointly analyze both static and dynamic features of malware, enabling the detection of complex behavioral patterns. To address the challenge of limited labeled datasets, a Generative Adversarial Network (GAN) is employed to generate synthetic data, thereby enriching the training samples and improving model diversity. The framework incorporates a data processing pipeline, including feature engineering, balancing, and real-time traffic analysis. To enhance transparency and trust, Explainable AI techniques such as LIME and SHAP are employed to interpret feature contributions and model decisions. Furthermore, a Flask-based web application is developed to enable real-time malware prediction through a user-friendly interface
The land registry administration is vulnerable to threats such as manipulation of centralized databases, lack of ownership proof, and synchronization fraud in transactions. In this paper, a decentralized approach is provided based on Ethereum blockchain, which involves role-based access control mechanisms between Administrators, Registrars, Users, and Banks. The approach reduces on-chain memory usage through an off-chain data pinning process using Pinata IPFS by hashing physical deeds to 46-character Content Identifiers (CIDs) so as to ensure an immutable timeline of Encumbrance Certificates (EC). In order to prevent forging of unregistered entry, a Two-Step Land Verification Pipeline is used, which comprises the user's property claim and registrar's verification gate asynchronously. Moreover, a Three-Way Financing Engine clears out land purchases through a bundled title transferring and bank's collateral locking operation on one block only. Our approach has been successfully implemented as a full-stack using React.js, ethers.js and MetaMask wallet signature and validated in terms of optimal cryptographic data integrity and security against any transaction closing manipulation.
ABSTRACT: Speech is one of the most natural and efficient forms of communication among humans, and it is becoming increasingly integral to human-computer interaction in modern applications such as virtual assistants, smart devices, and automated customer service systems. Among the various technologies that leverage speech, speaker recognition stands out as a biometric method that identifies or verifies individuals based on the unique characteristics of their voice. This approach has gained significant traction due to its non-intrusive nature, ease of integration, and applicability across various domains such as security, forensics, and personalised user experiences. In this paper, we introduce a novel speaker recognition system specifically designed to identify speakers based on utterances in the Marathi language, a linguistically rich and widely spoken regional language in India. The system utilises Mel-Frequency Cepstral Coefficients (MFCCS) to extract distinguishing vocal features that closely mimic the human auditory perception. MFCCS are particularly effective in capturing speech’s phonetic and acoustic properties, making them a preferred choice in speech and speaker recognition tasks. Vector quantisation (VQ) is applied to optimise the feature set and reduce computational complexity. VQ compresses the high-dimensional MFCC feature vectors into representative clusters without significantly compromising accuracy, thereby enhancing system efficiency. These features are then processed using hidden Markov Models (HMMS), which are adept at modelling temporal sequences and dynamic variations in speech. HMMS offer a statistical framework that effectively captures the sequential nature of speech patterns, leading to more reliable speaker modelling and recognition.
