8333 research outputs found
Sort by
A Deep Learning Filtration Framework to Eliminate Not Safe For Work content in Digital media
No full textNot Suitable for Work (NSFW) is a term which is used on the internet to warn users about content with inappropriate material such as nudity which might not be suitable to access at any work or study environment.
Most websites and chat applications have functionalities of NSFW content filtration which can disable the uploading of images or texts containing explicit words. However, this technique is not error free as there are cases where the filter blocks even the safe content. Besides that, there is no process currently in action which can filter for video content. The challenge here would be implementation of NSFW filter in a video content to identify NSFW content hidden within a video, which normally cannot be deduced by filename or the video thumbnail. This research proposes a Deep Learning Filtration Framework to filter the NSFW content in a Video. The proposed framework combines a Detection Model and Filtration technique to extract the content out of video. The Deep Learning Model is implemented using YOLO (You Only Look Once) trained on NudeNet classifier dataset for training and LSPD Dataset consisting of 500,000 images and 4,000 videos, with 93,810 labelled instances in 50,212 images for testing. The model is evaluated based on accuracy of the prediction. Results demonstrate values for safe and unsafe image where in values more than 75 percent in confidence are considered NSFW.
This research is on interest to provide a safe environment in the social media and chat sites to help avoid accidental exposure to such content and help the social websites/applications sustain their main purpose of social interaction based on demographics and not allow unmoderated content to be distributed at ease
Safeguarding Sensitive Data - Detection in Unstructured Text Using Cutting-Edge Transformer Architectures
No full textThe detection of PII in unstructured text enables organizations to protect privacy and meet legal requirements for data protection in accordance with GDPR, HIPAA and CCPA. Often ordinary prescriptive methods fails while working on the complexities that appear with unstructured data that require enhanced approaches. This research focused on using transformer-based models including DeBERTa, RoBERTa, DistilBERT, Longformer, in enhancing NER methods intended for identifying PII. The present analysis was created using ‘Learning Agency Lab - PII Data Detection’ dataset available on Kaggle. these models were trained to detect different form of PIIs but not limited to names, email addresses and phone numbers. In these models, DeBERTa showed the best performance with an F1-score of 0.91 indicating high levels of precision and recall for all classes. Longformer was really promising for long texts because of its ability to maintain the context, while RoBERTa demonstrated a fairly reasonable balance between speed and accuracy. However, for certain rare PII types, including emails and identification numbers, it became challenging for all the models to hit the intended performances no matter the level of dataset balancing and augmentation. Hyperparameter tuning and dropout regularization were among other techniques that further enhanced models, increasing generalization and reduce overfitting. Limitations aside, class imbalance and inherent sparsity in certain PIIs, findings underlined potential of transformer-based models. Future research may explore better data augmentation techniques, boosting models with other methods, and domain-specific pretraining approach. Findings of this research are valuable for academic and industrial purpose to build large-scale efficient PP systems
DeepDefend: Optimized Multi-Model Approach for Network Intrusion Detection Using Deep Learning and IoT Security Enhancement
No full textNetwork intrusion detection has become an important component of current cybersecurity techniques due to the growing frequency of cyberattacks in both conventional and Internet of Things-based network environments. However, current intrusion detection systems (IDS) frequently have high false-positive rates due to limitations in accuracy, adaptability, and the capacity to detect minority attack classes. In order to overcome these obstacles, this study suggests a new hybrid framework that combines advanced deep learning (DL) models with conventional machine learning (ML) techniques. The suggested method ensures strong generalization and adaptability across various network environments by utilizing a variety of datasets, such as UNSW-NB15, NSL-KDD, Cyber Intrusion and ToN IoT. The Perceptual Pigeon Galvanized Optimization (PPGO) method which is especially used to optimize Long Short-Term Memory (LSTM) models for improved IoT security, lies at the core of the suggested solution. In order to achieve balanced performance across a variety of attack vectors, this structure places to strong focus on lowering false-positives rates while also greatly increasing the detection accuracy of minority attack classes. Significant improvements in detection accuracy of upto 98.5% on UNSW-NB15 and 97.8% on NSK-KDD datasets, false-positive reduction, and flexibility to different network configurations are shown by exploratory results from experiments. The results help in the creation of a scalable and robust intrusion detection system that addresses the constantly evolving and complex nature of modern cybersecurity risks. This study has the ability to greatly improve network security across business and IoT-based systems by offering an integrated and efficient solution, paving the way to more secure digital ecosystems
Transforming the Performance of Airline Industry Through Sentiment Analysis
No full textThe rapid growth in technology-oriented businesses have encouraged different industries to adopt modern approaches such as sentiment analysis for better understanding of their customers to help gauge their feelings regarding the provided service. This research paper focuses on gathering meaningful insights from airline company customers through sentiment analysis which can help transform the overall performance within the airline industry. This research employs machine learning techniques such as Random Forest and Naïve Bayes to critically assess customer sentiments based on airline company’s dataset sourced through Kaggle. The used dataset in this research focuses on several aspects of airline services offered throughout the journey and provides customer ratings on its key factors. The models were evaluated through metrics such as accuracy, precision, recall, ROC curve and AUC score. Random Forest outperformed Naïve Bayes with an accuracy of 89.1% and an AUC of 95.1% compared to 79.3% accuracy rate and 89.3% AUC of Naïve Bayes. Correlation through weights highlighted the key factors that airline industry must focus on to transform the performance and enhance user experience. In-flight entertainment, ease of online booking and online customer support were the key factors that derive most customer satisfaction. This research highlighted actionable insights for the airline industry through the effective deployment of machine learning techniques in sentiment analysis. These insights can help airline companies in enhancing user experience, retaining customer loyalty and create a competitive advantage to capture market share and overall transform the performance of their business
Optimising Direct Marketing Through Data-Driven Analytics and Predictive Models
No full textDirect marketing campaigns are essential for preserving and growing market leadership in today's demanding environment. This research focuses on enhancing the effectiveness of these campaigns through data-driven analytics, employing advanced methodologies such as exploratory data analysis (EDA), customer segmentation via K-means clustering, and predictive classification models including logistic regression, decision trees, and K-Nearest Neighbours (KNN). The study aims to optimise marketing campaigns by identifying profitable customer segments and accurately predicting customer responses. The research employs the CRISP-DM methodology to methodically address business objectives, prepare data, build models, and evaluate their performance using a dataset from iFood, Brazil’s leading food delivery service. The findings offer practical insights that help direct marketing campaigns, improve client interaction, reduce costs, and reverse profit declines. Beyond resolving business challenges, this study contributes to the theoretical knowledge of data-driven marketing strategies and provides insightful information for both academic research and real-world implementations in the industry
Machine Learning-Based Improved Cold Start Latency Prediction Framework in Serverless Computing
No full textIn the contemporary world of cloud computing, serverless computing is the game-changer paradigm that enables on-demand, scalable, and cost-effective ways to deploy applications. One such critical problem related to performance and user experience in serverless computing is cold start latency. This paper introduces a machine learning framework for the prediction and improvement of cold start latency with the objective of enhancing serverless computing efficiency. We develop predictive models for cold start anticipation and duration prediction using machine learning techniques based on historical data analysis, together with real-time metrics. Machine-learning-based regression, decision trees, and neural networks have to be developed addressing patterns and correlating traditional techniques that are incapable of doing this task. We show that the built predictive models are able to forecast cold starts accurately, thus enabling strategies for resource allocation and optimization in a proactive manner. This not only helps in making applications more responsive but also gives a fillip to efficiency in resource utilization an important factor in reining costs for the service provider. Also, we design the framework to adapt and scale to the dynamic nature of serverless environments. Through integration with cloud infrastructure, it seamlessly augments the current state-of-the-art in offerings for serverless. This research fills an important gap in performance through effective means for the practical elevation of deployment and execution of serverless applications. This machine learning approach therefore represents a substantial step forward in the optimization of serverless computing for more effective and efficient cloud services
Minimizing Cold Starts in Serverless Environments with Predictive Optimization Approach Using Bi-LSTM and Genetic Algorithms
No full textServerless computing has revolutionized cloud applications because it only focuses on the server as an application service. However, the ”cold start” problem, where functions experience latency that is significantly higher when these functions have been inactive for a while, remains a critical issue when it comes to function performance. Prior work has suggested various solutions, such as static pre-warming techniques and partial runtime environments. Although these methods bring enhancements, they lead to the expenditure of more resources and are unsuitable for varying workloads. These shortcomings are mitigated in the current research by employing Bidirectional Long Short-Term Memory (Bi-LSTM) neural networks and Genetic Algorithms for a predictive optimization strategy. The Bi-LSTM model forecasts future cold start events following prior invocation probabilities, while GA enhances pre-warming strategies during runtime to reduce latency and resource consumption. This approach is different from previous methods in that it provides scalability and performance-optimized solutions, as required by the workload in real time. The solution used AWS Lambda in a serverless framework where performance was assessed relative to accuracy, precision, recall, and F1-score. The results prove that there are substantial enhancements in the field of cold start latency as well as application performance that testify to the reliability of the proposed predictive optimization strategy. The present research makes an innovative contribution by developing a model and using optimization methods that can be easily implemented and are cost-optimized to solve a well-known problem in the context of serverless computing
Enhancing Load Balancing Efficiency In Dynamic Workload Environments Using Enhanced Genetic Algorithm and Machine Learning
No full textCloud computing has gained significant popularity in recent times, providing scalable and elastic IT-enabled resources ’as a service’ to customers. There still exists issues with load balancing in cloud computing while handling dynamically dependent workloads. The improper distribution of loads in the cloud environment ultimately results in depreciation of overall system performance. To achieve this, multiple studies utilising various methods of both Enhanced Genetic Algorithm (EGA) and Machine Learning (ML) have been carried out. These methods often lack the vision and execution of integrating both technologies and consider the outcomes of this integration. This study aims to achieve more effective load balancing in cloud computing networks using an Enhanced Genetic Algorithm (EGA) combined with multiple Machine Learning (ML) targeted for the real-time load balancing problems.
An algorithm that has the capabilities to handle dynamic workloads and is equipped with advanced technologies that make use of predictive analysis can act as research area and a concept which can potentially resolve the issues with current present algorithms. This research proposes an EGA with Selection, Crossover and Mutation techniques in a virtual simulation of a Datacenter with Virtual Machines (VMs) that is capable of providing outputs as Resource Utilisation and Execution Time. The algorithm was executed for ‘n’ iterations to gather dataset for ML model training. The various Machine Learning models results showcased Linear Regression outperforming other techniques to achieve the objective
Empirical Study of Cloud Deployment Strategies: Guiding the choice between Containerization, Traditional and Hybrid Deployment
No full textThe empirical study provides a comprehensive evaluation of cloud deployment strategies – containerization, traditional virtual machines (VMs), and hybrid methods for three application types like static web applications, database web application and multithreaded applications with RabbitMQ. Motivated by the need for practical, data-driven guidance for cloud practitioners, the study evaluates key metrics such as performance, scalability, cost, reliability, and operational complexity. The findings shows that containerized deployment offer better performance and scalability for static web applications, hybrid deployments excel in performance, scalability and reliability for database web applications and multithreaded applications but both deployment strategies require complex setups which increases the operational complexity. While traditional VM deployments offer easy setup and low-cost offering usability for smaller applications and applications which do not have much load like academic projects or proof of concepts. A decision tree-based recommendation tool was developed to support practitioners in selecting appropriate deployment strategies based on the empirical data. Despite some of the limitations, including short evaluation period and resource constraints on scalability tests, this study shows a direction for future research in long performance analysis, broader application types, in depth scalability test and enhancing the recommendation tool. This future work will also help in commercializing this research study by the support of recommendation tool. The research ultimately provides actionable insights and practical tools to optimize cloud deployment strategies for its users, to ensure informed decision-making based on application requirements and scenarios
Hybrid models Cloud-based Enhancements for Air Quality Prediction Systems
No full textThis paper analyzes the creation and implementation of a hybrid model for air quality forecasting that incorporates statistical and machine learning techniques within the cloud computing paradigm. A major goal was to use the SARIMA (Seasonal AutoRegressive Integrated Moving Average) and Random Forest models with the aim of increasing predictive accuracy and reliability. First, LSTM could not be introduced to the initial trials in Cloud9 as the tool didn’t have TensorFlow - a prerequisite for running and integrating LSTMs into one model. Key results thus conclude that the SARIMA combined with Random Forest has indeed been deployed in the cloud, as the program developed in Google Colab proved successful. The deployment process involved linking the application to GitHub, which triggered a pipeline that facilitated the deployment to Elastic Beanstalk. Such a process underscores the role of cloud computing in dealing with large datasets and enabling real-time data processing for air quality predictions. The study points out the critical role of efficient model integration and cloud infrastructure in dealing with environmental challenges. This research contributes to the field of environmental monitoring by demonstrating that traditional statistical models can be effectively combined with machine learning techniques. The results suggest that such hybrid approaches can significantly improve decision-making related to public health and environmental policy, ultimately fostering more sustainable urban development practices