Ph.D Theses

Permanent URI for this collection

https://ir.avinuty.ac.in/handle/avu/9

Browse

Now showing 1 - 10 of 10

A Framework for Developing an Enhanced Convolutional Neural Network Based Ensemble Learning Model for Alzheimers Disease Classification Using MRI Brain Images
(Avinashilingam, 2025-03) Chithra S; Guide - Dr. R. Vijayabhanu
The integration of machine learning techniques in imaging domain is experiencing a deep transformation. It enables systems to analyze massive amount of data, distinguish patterns, and make forecasts with minimal human intervention. Machine Learning is applied to various domains in healthcare sector like disease diagnostics, treatment planning, drug detection and patient management. The machine learning models impact the complex and data exhaustive fields like oncology, cardiology, and neurology. In medical imaging machine learning models can examine MRIs, X-rays to identify irregularities like lumps, fractures, or organ deformities with high accuracy, regularly beating the capabilities of human clinicians. This present study focuses on the brain neuron images in classifying the Alzheimer’s Disease (AD) stages, which aids neurologists to understand complex changes in the brain. Through brain imaging analysis, the study strives to diagnose AD in its premature stages. AD is a deteriorating brain ailment caused by brain cells degeneration that impairs memory and intellectual damage that disturbs lots of old age individuals across the globe. Its a permanent brain ailment that steadily wear away thinking and memory skills which finally disturbs even the basic tasks. The memory and cognitive functions are affected in AD which is the reason for dementia in older population. These computational techniques use algorithms to analyze the brain images to classify patterns and features related to AD. To evaluate medical pictures and to discover neurological ailments like AD computational techniques like Machine Learning (ML) and Deep Learning (DL) techniques are applied in recent times. The objectives are to develop the classification potential of AD stages using ML and DL methods derived from ensemble classification framework. The contributions of this research work primarily focus on the preprocessing framework to eradicate the noises in the brain neuron MRIs by applying various de-noising filters to enhance the image deviations and to attain upgraded classification performance. The segmentation process is performed to for skull removal from the brain neuron image by applying thresholding methods to obtain a perfect image of the brain structure. To address the imbalance problems, a transfer learning approach is used for feature extraction. The first layer is transmitted, followed by the retrieval of features from Convolutional Neural Network using AlexNet model for feature retrieval. Lastly, classification is achieved from the extracted features using ML algorithms such as Decision Tree (DT), K-Nearest neighbor (KNN), Support Vector Machine (SVM), Neuro Evolution of Augmenting Topologies (NEAT), and BAGGING for AD stage classification. This study proposed two hybrid classification techniques like BAGGING_SVM and BAGGING_NEAT. The first hybrid classification technique combines BAGGING and SVM approaches to classify brain neuron images. The second hybrid classification technique combines the BAGGING and NEAT approaches to classify brain neuron images.
An Optimized Convolutional Neural Network Based Ensemble Classification and Regression Framework for Classifying the Stages of Diabetic Retinopathy
(Avinashilingam, 2024-05) Valarmathi S; Dr. R. Vijayabhanu
Deep learning (DL) techniques provide optimized solutions in a wide range of applications, such as natural language processing, face recognition, speech recognition, image analysis, and much more. Deep learning progresses from machine learning models, where the learning data is associated with task-based methods. Deep learning is identified as an effective way to handle complex image representation. Recently, the insights gained from deep learning techniques have aided the healthcare industry, especially in the medical imaging sector. Medical imaging is one of the high-priority areas for potential research with computer-aided medical devices, especially for disease diagnosis, disease monitoring and treatment. Internal organs such as the brain, retina, lungs, abdomen, kidneys, and much more can be captured in detail using medical imaging technology. This study focuses on exploring retinal disorders, which aids ophthalmologists in identifying the stages of diabetic retinopathy disease. Diabetic Retinopathy (DR) is an eye disease that affects the vision of a diabetic patient and can lead to blindness in its advanced stages. The rising number of diabetic patients worldwide is a necessity for emerging techniques in the present era. Scanning the retinal image to analyze the blood vessel layers at the rear of the eye is performed in retinal biometrics. The seepage on blood vessels in the retina in diabetic patients is the cause of permanent blindness. A digital photograph of a retina is used for screening patients with DR and Glaucoma diseases. Deep learning models aid in the classification of retinal images, providing optimized solutions. The objective of this study is to improve the classification performance of diabetic retinopathy stages using an optimized convolutional neural network-based ensemble classification and regression framework. A deep learning technique, Convolutional Neural Networks (CNNs), is employed in the form of pre-trained resnet-34 for DR stage classification. The contributions of this research work primarily focus on the preprocessing and augmentation phase, where a two-stage image preprocessing framework is proposed that involves a wavelet-based hybrid denoising method to eliminate Gaussian and Salt-and-pepper noises present in retinal fundus images followed by contrast enhancement, and augmentation to balance the dataset classes. As a part of feature extraction and classification, three different CNN-based deep learning models have been developed. These models aim to improve the performance of DR multi-class classification. The contributions in this section include, 1. Application of the Multi-Scale Attention (MSA) mechanism to a pre-trained CNN model, ResNet-34 (Residual Neural Network), in combination with a gradient-boosting classifier for DR classification tasks. 2. The utilization of Special Generative Adversarial Networks (SGAN) to generate realistic retinal images is followed by ensemble classification and regression blocks, as well as a Multilayer Perceptron (MLP) classifier for DR stage classification tasks. 3. Mine Blast Algorithm (MBA) enhancement to select the optimal set of hyper-parameters for tuning the deep learning model, thereby improving the classification performance of DR stages.
Cervical Cancer Detection and Classification in Pap smear Images using Enhanced Deep Learning Models
(Avinashilingam, 2024-11) Soumya Haridas; Dr. T. Jayamalar
Cervical cancer (CC) is the most significant contagious disease possessing women’s health by infecting Human PapillomaVirus (HPV) in cervix. Considering, the life daring outcomes of cervical cancer in later stage, early detection is considered crucial. However, past studies employed manual methods like Manual Liquid-based Cytology (MLBC) and Visual Inspection with Acetic acid (VIA) to identify cancerous cells. Meanwhile, the promising limitations including a high error rate, labor-intensive processes, and the need for specialized expertise have been witnessed in existing studies. Furthermore, Artificial Intelligence (AI)-based solutions are explored in this study to overcome the above mentioned shortcomings. Since, AI models are capable of analyzing huge volume of datasets to achieve precise results, it showcase more accurate detection and classification of cervical cancer cells. At present, the AI-based solution for cervical cancer detection and classification has reported suboptimal accuracy in their models. The major aim of this research is to enhance the accuracy of AI-based solution for CC detection using enhanced Deep Learning (DL) models. Three DL models have been enhanced using dissimilar pre-processing and segmentation technique with three distinct mechanisms for accurate classification of cervical cancer cells. These models are evaluated using two datasets: the Herlev dataset used for classification of single-cell images and SIPaKMeD for multi-cell classification in cervical cancer. The methodology encompasses four key stages in three models for detection and classification of cervical cancer: pre-processing, segmentation, feature extraction, and classification. The three models for CC detection and classification are described below. In Model-1, pre-processing based on diffusion stop function using Contrast Limited Adaptive Histogram Equalization (CLAHE) is utilized along with Topographic Weibull bounding-based segmentation, and segmented image is trained using radiance and variance enabled Deep Learning Neural Network for detection and classification. Model-2 pre-processes by employing the combination of Anisotropic Diffusion Filter (ADF) – histogram-based pre-processing and improved-Weighted Fuzzy C-Means (i-WFCM) - based segmentation; the CC is detected and classified using Restricted Boltzmann machine –Deep Belief Network (RBM-DBN). Model-3 performs ADF- Dragon Fly Optimization-based pre-processing along i-WFCM with Grasshopper Optimization Algorithm (GOA) -based segmentation techniques; Further, Deep Convolutional Neural Network with Rectified Linear Unit (DCNN with ReLU) is incorporated for classification. These models classify the single cell images from Herlev dataset into 7 classes; superficial, intermediate, columnar, light, moderate, severe, and carcinoma. Whereas multi-cell images from SiPaKMed dataset into 5 classes; superficial, koliocytotic, parabasal, dyskeratotic, and metaplastic. In addition, the proposed three models are evaluated using different performance measures namely accuracy, precision, recall and F1-measure. Out of the three models Model-3 achieves better performance than other models with 97.2% accuracy, 91.3 % precison, 96.9 % recall, and 94.02 % F- measue for multi-cell Classification; in contrast, model-3 achieved 96.7% accuracy, 85.1 % precision, 95.2 % recall, and 89.8 % F-measure for single cell classification in cervical cancer detection. However, evaluating the present research work using real images in a software application could aid the medical professionals in real-time for identifying the cancerous cells with the aim of saving patients’ lives lynching in cervical cancer disease.
Complexity Aware Intelligent Intrusion Detection for Ddos Attacks
(Avinashilingam, 2025-01) Kalaivani M; Guide - Dr. G. Padmavathi
Distributed Denial of Service (DDoS) attacks pose a major risk to the availability and security of modern network infrastructures. Their growing complexity and scale have outgrown traditional defense methods. Current solutions such as firewalls and standard intrusion detection systems often can't adapt to handle complex and changing intrusion patterns leading to inefficiencies in detection and mitigation. This issue majorly affects industries like finance and e-commerce where security breaches can cause huge damage. To address these problems, this study suggests a new smart Intelligent Intrusion Detection System (IDS) framework that understands complexity. This aims to detect threats with better accuracy, minimized computing power, and have few false alarms. This helps to boost security and availability against the changing world of cyber threats. This thesis aims to create a complexity aware intelligent IDS to fix the problems with current systems. It combines cutting-edge machine learning (ML) and deep learning (DL) models with nature-inspired optimization algorithms to make DDoS attack detection more accurate faster, and stronger. Feature Engineering is the major focus in identifying the right features and making the intrusion detection model better with minimized resources. The novelty of the research lies in developing advanced, complexity-aware intrusion detection systems for DDoS attacks, leveraging innovative methods like Combined Filter for Feature Selection (CFFS), bio-inspired Dragonfly Optimization, Panthera Leo Optimization, and an Attention-Enabled Gated Recurrent Network (AEGRN) to achieve significant detection accuracy, computational efficiency, and adaptability across diverse datasets. A significant contribution of this research is the development of four distinct methodologies. The first contribution enhances the detection of single-vector DDoS attacks using a Combined Filter for Feature Selection (CFFS) integrated with a Decision Tree (DT) classifier. This method achieved an accuracy of 97.69%, with precision and recall exceeding 99% and a false positive rate of 6.32%. However, its performance declined when applied to multiple flooding attacks, indicating the need for more robust techniques. The second contribution introduces the Improved Dragonfly Optimization Algorithm (IDOA) alongside a Decision Tree (DT) classifier to enhance detection accuracy for multi-vector DDoS attacks. This approach achieved 98.89% accuracy, with precision and recall above 97%, an F-score of 98%, demonstrating significant efficiency while leaving room for further improvements in accuracy and efficiency. The third contribution involves an Integrated Intrusion Detection System (IDS) based on the Panthera Leo Optimization (PLO) technique combined with a multilayer feedforward network. This method successfully managed network traffic complexity and variability while maintaining low computational latency. Using the CICDDoS2019 dataset, it achieved a prediction accuracy of 96.8%. The final contribution presents a novel Attention-Enabled Gated Recurrent Network (AEGRN) for detecting DDoS attacks across multiple datasets. This IDS demonstrated over 98% generalization accuracy across various datasets, with an average processing time of 17.4 seconds per epoch. Self-attention maps with BiGRU and feedforward networks proved beneficial in achieving better classification accuracy with reduced complexity and processing time. The proposed models have been evaluated using standard performance metrics such as accuracy, precision, recall, F1-score, and computational time. Statistical validation using techniques such as ANOVA and p-tests has confirmed the reliability and significance of the improvements observed. This thesis provides a novel and effective framework for detecting DDoS attacks through the integration of advanced ML, DL, and optimization techniques. The proposed solutions demonstrate notable performance in terms of accuracy, scalability, and computational efficiency, making them suitable for deployment in real-world scenarios. Future research should focus on validating the effectiveness of the developed model on real-time datasets to better reflect real-world cyber threats. Additionally, efforts should be made to assess the model’s capability in identifying and mitigating AI-enhanced and Deep DDoS threats, ensuring robustness against evolving attack strategies that leverage artificial intelligence.
Deep Learning-Based Facial Expression Recognition for Analysing Learner Engagement in Mulsemedia Enhanced Teaching
(Avinashilingam, 2024-11) Mohana M; Dr. P. Subashini
In the current digital era, technology-enhanced learning is evolving rapidly, setting new trends in educational environments and enabling students to learn more efficiently than ever before. Conventional learning course content typically involves only two sensory modalities—audio and video—which limits its ability to engage learners fully. In contrast, immersive learning course content and environments incorporate multiple senses, allowing learners to interact with multimedia content in ways that go beyond sight and sound. This approach, known as mulsemedia (multiple sensorial media), posits that engaging sensory channels—such as audio, visual, haptic, olfactory, temperature, gustatory, and even airflow— can significantly reinforce the learning process. Furthermore, measuring learner engagement is essential to ensuring that learners remain actively involved in learning. Various detection methods can assess engagement levels; in this study, we focus on analyzing engagement through facial expressions, particularly in a mulsemedia-synchronized learning environment. Modern Facial Expression Recognition (FER) systems have achieved significant results through deep learning techniques. However, existing FER systems face two primary challenges: overfitting due to limited training datasets, and additional complications unrelated to expressions, such as occlusion, pose variations, and illumination changes. To improve the performance of FER in analyzing learners' engagement within a mulsemedia-based learning environment and to address some of these challenges, we propose three key aspects. In our first study, face detection is a crucial step for identifying and cropping faces to train FER models. We observed that the conventional Viola-Jones face detection algorithm often produced false positives, particularly in complex images containing multiple faces or cluttered backgrounds. To address this issue, we enhanced the Viola-Jones algorithm by integrating particle swarm optimization to improve prediction accuracy in challenging images. The integration optimizes threshold selection and refines feature selection, enabling AdaBoost within the Viola-Jones framework to focus on the most relevant features for constructing a robust classifier. This enhancement significantly reduces false positives by fine-tuning feature selection and cascade thresholds, thereby improving prediction accuracy in complex environments. In our second study, we observed that existing supervised FER approaches are inadequate for analyzing spatiotemporal features in real-time environments involving dynamic facial movements. To overcome this limitation, we introduced a fusion of convolutional neural networks and Bidirectional Long Short-Term Memory (Bi-LSTM) networks to recognize emotions from facial expressions and capture relationships between sequences of expressions. Our approach employs a VGG-19 architecture with optimized hyperparameters and TimeDistributed layers to independently extract spatial features from each frame within a sequence. These spatial features are subsequently fed into a Bi-LSTM, which captures temporal relationships across frames in both forward and backward directions. This fusion enhances the model’s ability to recognize emotions from expression sequences. The proposed method achieves significant accuracy in FER analysis, with results compared against baseline techniques. In our third study, we introduced a Deep Semi-Supervised Convolutional Sparse Autoencoder to address the limitations of supervised FER approaches, particularly their reliance on extensive datasets and the challenges posed by imbalanced facial expression distributions, which can adversely affect model performance. This approach consists of two main stages. In the first stage, a deep convolutional sparse autoencoder is trained on unlabeled facial expression samples. Sparsity is introduced in the convolutional block through penalty terms, encouraging the model to focus on extracting the most relevant features for latent space representation. In the second stage, the trained encoder’s feature map is connected to a fully connected layer with a softmax activation function for fine-tuning, forming a semi- supervised learning framework. This approach enhances FER accuracy in real-time environments. Furthermore, these two approaches were conducted using the Extended Cohn- Kanade+, Japanese Female Facial Expression, and an In-house dataset. Model performances are evaluated using metrics including accuracy, precision, recall, F1-score, the confusion matrix, and the receiver operating characteristic curve. Finally, all proposed methods were integrated to effectively analyze learner engagement levels in mulsemedia-synchronized learning environments. To achieve this, a mulsemedia-synchronized web portal was developed, incorporating olfactory, vibration, and airflow effects. The FER system mapped eight facial expressions to three engagement levels—highly engaged, engaged, and disengaged—based on the system’s predicted probability scores and predefined threshold values. The final results demonstrate that mulsemedia-based learning significantly improved learning outcomes and memory retention compared to conventional methods.
Modified Extreme Learning Machine Algorithm with Deterministic Weight Modification for Investment Decisions Based on Sentiment Analysis
(Avinashilingam, 2024-10) Kalaiselvi K; Dr. Vasantha Kalyani David
The trading of stocks contributes to the growth of the commodity economy by driving a significant quantity of capital into the stock market, which improves the organic configuration of corporate capital through capital concentration. Consequently, the stock market is seen as a measure of the financial activity of a nation or area. Specifically, since it can precisely depict the supply dynamics of the stock market, the trading price of the stock frequently acts as a measure of the price and quantity of the stock. Timely and precise stock price prediction and analysis are essential for both investor decision-making and the constancy of the national economy by increasing returns and decreasing risks. Consequently, researching stock projections can help depositors make wise decisions that will advance society and yield rewards for themselves. The intricacy of financial time series presents challenges that ML can handle with its strong data processing skills. Consequently, there are a lot of opportunities for ML and finance together, but there hasn't been enough research done in this field. Furthermore, the stock market is not entirely objective and does not always follow scientific principles due to humans' emotional, psychological, and behavioral traits. Recent studies have also demonstrated that investor sentiment may play a significant influence in stock market investing.The present study proposed a modified extreme learning machine (ELM) algorithm with deterministic weight adjustment to increase the precision and dependability of sentiment analysis-based investment decision-making. To capture investor mood, the approach incorporates financial sentiment research from news articles, social media, and market patterns. With deterministic weight initialization (DWM), the ELM algorithm achieves more consistent model performance than standard ELM techniques that use random weight initialization. The suggested model is a potent tool for sentiment-driven investing strategies since it shows improved prediction accuracy, quicker learning, and robustness in financial forecasting.
Optimized Hybrid Forecasting Model using Enhanced Deep Learning Based Approach for Transaction Management in Commodity Trading
(2024-11) Barani Shaju; Dr. N.Valliammal
Forecasting commodity price movements are critical for economics and trade decisions for industries that rely on raw materials, agricultural goods, and natural resources. Commodities include oil and gas to agricultural produce, metals, and minerals. In trade market, commodity price forecasting and trend analysis are critical for making informed decisions thereby reducing risks and maximizing resource allocation. Forecasting commodity prices on the other hand is a difficult endeavor riddled with challenges like market volatility, data quality and availability, short and long-term forecasts, environmental issues and external factors like as government rules, trade agreements, currency exchange rates are few to mention. Transaction management in commodity trading operates based on trading strategies and traded Volume (V), considering price per lot size. The effectiveness of transaction management is largely determined by the ability to deliver consistently high returns while adhering to clearly defined and specific turnaround timelines, ensuring optimal outcomes and overall success. In the commodity trading market, traders develop and implement transaction management strategies based on the volume and lot size of each commodity investment, taking price into account. Research and analysis suggest that leveraging analytical methods can provide valuable market insights, enabling traders to maximize profits. Given the complexity and multidimensional nature of data, alongside current economic conditions, market trends, and fluctuations, employing Deep Learning (DL) techniques with time series decomposition is recommended for building effective forecasting models. Time series-based forecasting has gained significant traction as data-driven derivatives play a crucial role in decision-making. Time series represents a sequential collection of statistical data arranged in chronological order. Time series analysis involves predicting future data trends based on historical patterns observed over previous time periods. The focus of this study is to improve the advantages of commodity trading by increasing the efficiency and effectiveness of transaction management strategies.With a wide range of fluctuations due to market dynamics, trader finds it as a challenge to understand price movements and make trade decisions. This research primarily aims to develop an advanced analytical model capable of forecasting commodity trade prices with improved accuracy. The model is designed to effectively manage outliers and anomalies, offering reliable investment guidance to commodity traders based on predictive insights. The first phase of work aims to develop an efficient forecasting model to predict commodity trade prices, focusing on managing outliers to improve accuracy. The model leverages Long Short Term Memory (LSTM) with multiple-kernel configurations, within each layer to enrich representations and optimizes weights over the lower and upper bounds. Hyper-parameters are tuned using the Keras tuner for optimal performance. Dropout regularization is employed to reduce the risk of overfitting by probabilistically decreasing network capacity and regularizing the output. The model also uses dropout to regularize and optimizes weights over the span bound, which are validated using a "Train - Validate - Test" approach. This methodology ensures robust and accurate price forecasting for commodity trading through advanced DL techniques and thorough validation processes. In this second phase of research work, aim is to enhance the price forecasting period to wider horizon period ranging from one day to a month long period along with signal optimization to predict prices for commodity trading using a metal dataset from MultiCommodity Exchange of India (MCX). The prediction of a time-series is a challenging task because of the inherent complexity of data, non-linearity of time series information and significant noise. The criteria include detecting anomalies through white noise detection and applying min-max normalization for calculation of threshold limits. The technical choices involve using Neural Prophet (NP), exponential window slicing, and cross-validation with a "forward chaining" mechanism. The research outcome includes forecasting across the window period, providing market insights based on seasonality, improving error detection performance metrics and increasing optimized signal strengths. The model reveals seasonality and trend based on the time period forecast. In the third phase of research work, the objective is to build a highly accurate model for forecasting commodity prices using an integrated approach combining E-MKLSTM and Neural Prophet with regressors. The aim is to achieve this by incorporating outlier detection, regressor analysis, min max normalization and Generative Additive Model (GAM) to implement holiday components. The enhanced multikernel LSTM-NP model captures a wide range of temporal patterns and external influences, making predictions more accurate, robust, and interpretable. This hybrid model captures complex patterns and non-linear relationships within the data, by efficiently handling outliers and minimizing loss functions. Novelty of this approach revolves around the mechanics used to handle outliers and improvise signal strength, which enhances the robustness of the model. The impact of anomalies are thereby minimized which boosts the performance capability of the model. Learning derivative of this approach indicates that calculative tradeoff can be made by traders based on forecasted data. When decision making is backed based on the data insights, quality of decision making of the trader in response to market dynamics will yield long term benefits. This model brings significant optimization in forecasting commodity prices by effectively addressing outlier challenges and enhancing overall performance efficiency. The enhanced Multi-Kernel LSTM (E-MKLSTM) component excels in managing long-term dependencies, ensuring the retention of relevant information over extended periods. Furthermore, Neural Prophet's ability to model seasonality and trends enables it to accurately capture recurring patterns and long-term dynamics, which are essential for reliable forecasting. Context awareness is another key feature incorporated with inclusion of regressors, such as country-specific holidays for required region, thereby enhancing the model's ability to account for external factors influencing commodity prices. The model's robustness to non-stationary data components further demonstrates its suitability for real-world applications, where data properties often change over time. It is versatile, capable of handling both short- term and long-term forecasting needs, providing flexibility for various applications within commodity trading. With non-linear factors that impact trade, proposed research approach aims to forecast price movements and create future timeframes with seasonality after evading anomalies. Evading the outliers and white noise along with focus on seasonality brings ease of focus on the business implementation. This model is suggested as an optimized, reliable method suitable for complex, multi-faceted business domains with robustness and high interpretability in commodity price forecasting. Leveraging the combined strengths of advanced analytics, deep learning algorithms and time series decomposition, the research achieves remarkable accuracy.
Prediction of Heart Diseases Risk Using Novel Machine Learning Techniques
(Avinashilingam, 2024-11) Anuradha P; Guide - Dr. Vasantha Kalyani David
Heart diseases are a major health concern globally and early detection of the disease plays a crucial role in reducing mortality rates and improving patient outcomes. With the advancements in machine learning, there is an increasing focus on utilizing these techniques to enhance the prediction and diagnosis of heart diseases, even before symptoms manifest. Machine learning (ML) research in this field aims to develop accurate and efficient models that can assist in early detection, risk assessment, and clinical decision-making. There are a variety of ML methods making use of stand-alone classifiers and hybrids. However, the results from these models vary considerably between various cardiac datasets and/or they are not modeled using both low and high dimensional data. With increased dimensionality of data, it becomes imperative to address shortcomings of existing feature selection and prediction approaches in handling such datasets with complicated feature relationships and significant degrees of redundancy. This research identifies and addresses issues with existing feature selection and classification methods and proposes novel and improved feature selection and classification techniques towards enhancing and improving heart disease prediction performance. In the first stage of work, feature selection using Feature Importance (FI) ranking of Gradient Boosting algorithms is done and a significant reduction in the search space of feature subsets is identified. Next, this research work proposes a novel feature selection algorithm called ModifiedBoostARoota (MBAR), which identifies the risk parameters that strongly contributes to the prediction of heart disease. This algorithm incorporates CatBoost as the base model and utilizes a novel feature elimination process. In the second phase of the work, a novel Super Learner Ensemble Model (SLEM) is proposed to perform on features selected by MBAR. The SLEM model is an integration of diverse ML base models selected by repeated stratified k-fold cross validation. A meta learner logistic regression is employed to learn from the predictions of the base classifiers. By backward elimination method, an optimal combination of classifiers in SLEM was identified as Catboost and Decision Tree, in order to improve the classification time complexity and performance. The performance of the SLEM model improved when used on features selected by MBAR compared to its performance on datasets with no feature selection. In the third phase of the work, to further improve the classification by selecting an optimal combination of base classifiers in the Super Learner Ensemble Model (SLEM) model, a new Optimized Super Learner Ensemble Model (OSLEM) is proposed. OSLEM utilizes the Whale Optimization Algorithm and pairwise divergence measure to select an optimal base classifier combination. The performance of the final proposed model where ModifiedBoostARoota algorithm is used for feature selection and Optimized Super Learner Ensemble Model (OSLEM) is used for classification was analyzed on both low and high dimensional heart datasets. The proposed model presented high performance in terms of recision, recall, f1-score, specificity and accuracy, when compared with existing models across the heart datasets. The proposed model improved prediction accuracy while being robust against overfitting. Overall, the contributions from this research work would improve the accuracy, efficiency, and decision-making support of heart disease prediction systems, ultimately benefiting clinical practice.
Securing the VANET through a Hybrid Approach by Mitigating DoS Attacks and its types with Self-healing and Immunization
(Avinashilingam, 2025-01) Rama Mercy S; Guide - Dr. G. Padmavathi
Vehicular Ad Hoc Networks (VANETs), crucial for Intelligent Transportation Systems (ITS), face significant security threats, especially Denial of Service (DoS) and Distributed DoS (DDoS) attacks. These attacks disrupt communication, leading to packet loss, increased latency, and reduced reliability. While existing solutions like trust-based models, cryptographic techniques, and machine learning approaches exist, they often fall short in detection accuracy, energy efficiency, adaptability to mobile environments, and managing system overhead. This research, titled "Securing VANETs through a Hybrid Approach: Mitigating Denial of Service (DoS) Attacks and its types with Self-healing and Immunization," proposes a three-phase methodology to enhance the security of VANETs. It leverages a hybrid approach having six key contributions with the major objective to secure VANETs—a key part of Intelligent Transportation Systems—from DoS attacks by detecting and preventing these attacks including self-healing and immunization features. The scope of the research tends to focus on DoS attacks and its types with multi-layered defense incorporating security and robustness. In Phase 1, the objective is to detect and isolate vehicles under malicious DoS attacks with optimized feature selection using GLW-SLFN (Glow-worm Single Layer Feed Forward Neural Network). MCOD-LR (Micro Cluster Outlier Detection and Linear Regression) is applied to detect malicious behavior for multi-class DoS attacks. Furthermore, Kernel Density Estimation and Entropy-based SVM (Support Vector Machine), incorporating trust factors, are leveraged to detect, predict, and classify DoS attacks. A Bayesian aggregate model, in conjunction with Self-healing AIS (Artificial Immune Systems), ensures the continuous monitoring, detection, and isolation of these attacks.The changing topology in VANET remains a challenge in securing VANET operations. This challenge is addressed in phase 2, as the traffic signals are encrypted using Triple Random Hyperbolic Encryption (TRHE) integrated with Hex-Tuple Matched Mapping, which classifies twelve types of DoS attacks. The classification relies on mapping reports and a Deep Trust Factorization Neural Network (DT - NN).Furthermore, to achieve stable data transmission and routing even with dynamic network topology, phase 3 is proposed to immunize the behavior of its clusters by the Deep Trust Factorization Neural Network (which provided trust scores), the Moth Flame Optimization (MFO) Algorithm, and Cache Parallelized Circulation Link Routing (CCL). The system achieved stable data transmission and routing, even with dynamic network topology, due to the immunized behavior of its clusters. The Moth Flame Optimization (MFO) algorithm optimizes the Packet Delivery Rate (PDR) essential for ensuring data is delivered efficiently. This system efficiently creates stable clusters and identifies reliable relay nodes within a VANET. This feature enables the isolation of malicious nodes, directly leading to a significant increase in the Packet Delivery Rate (PDR). The performance of Phase 1 demonstrated significant improvements: a 37% increased detection rate over AODV, 32% over Trust-based methods, and 20% over Firecol. The approach also reduced energy consumption by 38% compared to AODV, Trust-based, and Firecol. Furthermore, it achieved a 25% lower latency, markedly outperforming AODV (95%), Firecol (58%), and the Trust-Based Framework (27%). Building on this, Phases 2 and 3 collectively enhanced overall performance, resulting in a minimized packet loss of 0.5 bits for 200 nodes, a maximized attack detection accuracy of 97%, and a Packet Delivery Ratio (PDR) of 98%. These figures represent a substantial improvement over existing techniques: Trilateral Trust (42% accuracy, 60% PDR), Host- based Intrusion Detection System (H-IDS) (60% accuracy, 70% PDR), Multi-filter (80% accuracy and PDR), and Stream Position Performance Analysis (SPPA) (90% accuracy and PDR). This approach demonstrates remarkable scalability and adaptability, particularly in challenging environments with high node mobility and dense vehicular traffic. The methods ensure resilient network operations in intelligent transportation systems, delivering reduced energy usage, lower communication delays, and high detection accuracy for secure, reliable, and scalable communication. This research provides highly relevant solutions for real-time VANET applications, effectively incorporating self-healing, immune-inspired mechanisms.
Specular Reflection Removal in Smart Colposcopy Images Using Deep Learning Models for Enhanced Grading of Cervical Cancer
(Avinashilingam, 2024-04) Jennyfer Susan M B; Dr. P. Subashini
Cervical cancer is a significant global health concern, ranking as the fourth most common cancer among women, primarily caused by Human Papillomavirus (HPV) affecting the lower uterus. Despite preventive measures like HPV vaccination and screening programs, many women hesitate due to invasiveness. Smart colposcopy, an advanced non-invasive approach, captures cervix images for examination. However, white specular reflections caused by body moisture pose challenges, hindering accurate analysis and potentially leading to misclassification of dysplasia regions. This research aims to improve cervical cancer grading by identifying and removing specular reflection from smart colposcopy images. Initial focus lies on specular reflection identification, employing RGB and XYZ color spaces for optimal detection. The proposed intensity-based threshold method accurately identifies specular reflection on XYZ color, overcoming challenges posed by vaginal discharge and acetowhite regions. In the second phase, pixel-wise segmentation models like Fully Convolutional Neural Network (FCN), SegNet, and UNet Model are employed. On comparison analysis of the segmentation model, the UNet model indeed demonstrates higher accuracy. However, when it comes to the intersection of Union, the UNet model falls short due to the overlapping of segmentation. To address this limitation, different versions of the UNet model are compared, and the UNet++ model emerges as the most promising, exhibiting higher intersection of union metrics. Subsequently, the UNet++ model is fine-tuned to optimize its performance in segmenting reflection regions. After segmentation of the reflection, the empty region should be filled with neighboring pixels to improve the quality of the images. A novel Bilateral-based Convolutional Inpainting model fills eliminated regions, improving image quality. This model outperforms traditional methods, particularly in medical image applications, showcasing efficacy across different masking ratios. Enhanced images, with removed specular reflection, undergo grading using DenseNet121, VGG19, and EfficientNet. Trained on both enhanced and non-enhanced images, classification models achieve significantly improved prediction accuracy with enhanced images, underscoring the enhancement technique's impact on cancer stage classification. This research offers valuable insights into medical image analysis, presenting an integrated approach for cervical cancer grading. The proposed methodologies exhibit promising results, laying the groundwork for further advancements in women's health and cancer diagnosis.

Browse

Browsing Ph.D Theses by Subject "Computer Science"

Results Per Page

Sort Options