8333 research outputs found
Sort by
A Hybrid HSV and YCrCb OpenCV-based Skin Tone Recognition Mechanism for Makeup Recommender Systems
No full textSkin detection technology serves multiple purposes across various sectors, including surveillance, criminal justice, and healthcare. This study focuses on skin detection by extracting RGB values of skin tones from facial images of diverse ethnic backgrounds. Utilizing a three-tier OpenCV-based architecture, the approach encompasses skin detection, skin tone identification, and is tested in an application for recommendation of the most fitting makeup foundation shade, brand, and product. The research evaluates three color-space models for their effectiveness in skin and skin tone detection: HSV (Hue, Saturation, Value) with Gaussian blur, HSV alone, and a combination of HSV and YCrCb enhanced with gamma correction and image segmentation. The precision of each color space method was evaluated by measuring the difference between the predicted RGB skin tone values and the actual RGB skin tone values, utilizing the Delta-E metric for comparison. The hybrid model combining HSV and YCrCb color spaces emerged as the most accurate, achieving the lowest Delta-E average value of 16.68, thereby surpassing the other methodologies
Exploring the Impact of Social Media on Mental Health and Well-being: A Multi-Dimensional Analysis
No full textLately, the popularity and adoption of social media have increased globally, leading to an inevitable part of daily communication and connection. However, the widespread use of social platforms introduces significant challenges to mental wellbeing. Ongoing state-of-the-art research demonstrates a robust correlation between extensive social media usage and the worsening of medical conditions such as depression and anxiety. These approaches often struggle because they fail to adequately model key health-related information and screen usage frequency in their analyses. Moreover, most of the existing methods do not fully leverage advanced data analytics and machine learning technologies to effectively model the influence of social media algorithms on different age groups. In this paper, we employ a multi-model analysis, utilizing techniques such as decision trees, random forests, support vector machines (SVMs), and convolutional neural networks (CNNs), to gain a deeper understanding of the intricate relationship between social media use and mental health. In particular, we present four case studies exploring demographic risk factors, broader health consequences and specific health behaviors, and depression predictors associated with the use of social media. These analyses provide a comprehensive view of social media’s multi-dimensional impact on mental health and offer critical understandings for designing targeted interventions and policies
Supporting parents to navigate the modern digital ecosystem for themselves and their children
No full textTraining Policy for Privacy-Preserving Logistic Regression in Federated Learning Environments
No full textLogistic Regression (LR) is a widely used statistical model for classification problems. However, its training and evaluation in a shared environment increase the possibility of information leaking. A federated LR reduces security issues by using only locally available data for training. In a Federated Learning (FL) environment, LR receives the coefficients of local models to create the federated LR model, which is then distributed to update the local models. The exchange process does not leak confidential information when LR coefficients are encrypted. Homomorphic Encryption (HE) allows the merging of local LR models with privacy preservation (HE-LR). This work presents a novel training policy to reduce the training time with only slightly decreased quality in an FL environment with HE. We analyze the accuracy and time of FL policies with HE-LR that progressively reduce the amount of training data and exchange the LR coefficients in a privacy-preserving manner. The results show that the proposed policy can speed up the training time between 12% and 69%, compared to the traditional FL approach, with an average decrease in accuracy of 1.79% and 1.95%
Network Intrusion Detection using Supervised and Deep Learning Machine Learning Algorithms
No full textThe expansion of computer networks has exacerbated worries about network security and resulted in a number of infiltration attempts. The confidentiality, integrity, and availability of data and systems are compromised by these attempts. The necessity to handle the growing risks of cyberattacks is highlighted by statistics showing an increase in the frequency of malware assaults and denial of service situations. The increase in network traffic, the complexity of Network Intrusion Detection Systems (NIDS), and the variety of protocols and data transferred via contemporary networks are the three main issues that worsen network security. The existing traditional approaches are unable to detect new types of attacks, thereby necessitating the need for more robust solutions. The goal of this research is to increase the effectiveness of machine learning and deep learning models, which include some of the most applied classification approaches, namely decision trees (DT), logistic regression (LR), naïve bayes (NB), convolutional neural networks (CNN) and recurrent neural networks (RNN). Additionally, I examined the models' performance in binary classification as well as the effects of feature significance selection and hyperparameter adjustment on the CICIDS 2017 and UNSW NB15 benchmark datasets. Based on the findings of the experiments, the optimized decision tree is the best model for a network intrusion detection system with accuracy, F1-score, and AUC score of 99.27%, 99.26%, and 99.27% respectively on the CICIDS 2017 dataset. On the UNSW NB15 dataset, the scores were 99.28% across all metrics. It outperforms other machine learning and deep learning classification techniques and underlines the superiority to traditional IDS
Enhancing Predictive Analytics through Machine Learning Models in Cloud Computing Environments
No full textMachine learning (ML) models integration within cloud computing environment can significantly improve the capabilities of predictive analytics in all industries. In this research, we investigate the challenges and opportunities of gaining deployment of machine learning models within cloud environments, especially those relating to performance, scalability and cost efficiency. In this project we investigate several cloud deployment strategies — use of cloud services such as Amazon Web Services (AWS), and apply efficient resource management practices such as auto-scaling and model parallelism. The research features key components of which are to choose viable cloud computing platforms, to optimize machine learning models for cloud deployment, and to secure and guarantee the privacy of data in a cloud setting. A performance evaluation of trained and deployed a Random Forest Regressor model using the Yahoo Finance dataset based on metrics of accuracy, inference time, and cost. This study has yielded actionable insights on best practices for exploiting cloud based ML solutions for scalable and low cost predictive analytics and addressing ethical concerns of dealing with data privacy and model transparency. This work extends the body of research on cloud-based machine learning applications with important lessons for practitioners aiming to leverage ML driven applications in cloud environment
Evaluation of GWO -PSO Algorithm using AWS Lambda
No full textThe novel approach to load balancing in Cloud Environment with the help of integrating the AWS services with optimized hybrid GWO-PSO algorithm. The main objective of the approach is to attain a efficient and scalable traffic managing system with the help of Grey Wolf Optimization (GWO) and Particle Swarm Optimization (PSO) algorithm. The hybrid algorithm is implemented on AWS Lambda. Lambda is known for it’s serverless architecture and scaling capability. For dynamic traffic distribution the Application load balancer is used and EC2 instances acts as the backend server.. The main advantages are resource utilization through efficient workload distribution across the instances. The success criteria considered for this approach are based on CPU utilization, minimal error rates and optimal response time. Comparison is performed between the traditional VM based approach and the hybrid GWO-PSO algorithm on AWS Lambda to analyze the valuable insights for the system performance .The approach guarantees the robust fault tolerance
Optimizing Cloud Power In An Open Radio Access Network Based on Subscriber Behavior
No full textThe fast adoption of 5G has significantly changed mobile communications, due to the high speed, very low latency, and the ability to serve billions of devices. However, this rapid expansion comes at the cost of increased energy consumption, caused by the need for more Radio Units (RU) deployment in high-traffic areas because of the use of higher-frequency radio waves, which offer shorter transmission ranges. This challenge has increased the computational load on cloud based Distributed units and Central units further compromising energy efficiency. Open Radio Access Network (O-RAN) architecture was a paradigm shift by introducing disaggregation, virtualization and open interfaces as those features enable more flexibility and interoperable networks. Additionally, O-RAN introduced RAN Intelligent Controllers which enables important features such as closed-loop control and AI-driven decision-making.
This thesis proposes a novel AI-driven solution to optimize energy consumption in O-RAN networks by predicting RU power states based on subscriber behavior. What makes our work different from the traditional approaches, static power management, is the use of AI to adapt in real-time to network changes and user mobility ensuring more efficient and responsive energy use in 5G networks. To achieve this goal a virtual network simulation was created and Markov process used to simulate realistic user mobility and traffic patterns across different regions and times. The generated dataset was used to train Random Forest Classifier in AWS SageMaker, resulting in predictive accuracy of 99.09%. Finally, the trained model will be deployed in a real-time cloud-based environment enabling prediction request and response through API Gateway and AWS Lambda integration
Analysing The Effectiveness of Machine Learning Algorithms in Intrusion Detection Systems for IoT Networks
No full textNetwork Intrusion Systems are very essential in managing networks and mainly ensure to keep their security information safe. Conventional approaches based on rule analysis have mostly failed to provide effective solutions for complicated attacks by learning from the outcomes of IDS implementing in complex environments like cloud computing and IoT. This work examines Random Forest, Support Vector Machine and KNearest Neighbour ML algorithms for IDS using the NSL- Knowledge Discovery in Databases. The analysis of these algorithms which considers the classification accuracy and estimations by using the measures of precision, recall, F1 scores and random forest that results the best outcome with an overall prediction accuracy followed by KNN with 97.45 Percent and SVM with 95.54 Percent. The findings of the study that discuss about the feature selection, data preprocessing and the model evaluation as critical areas that are important to the enhanced performance of IDS. Moreover problems like lack of balance of datasets, the problems that arise due to big data and high computational complexity in real-life networks are still open to be solved. To overcome these types of stated limitations the study which effectively suggests to use oversampling methods such as SMOTE, hyperparameter tuning and the integration of deep learning to improve the process of anomaly identification. For future studies it is recommended to work with efficient real type datasets, together with ensemble solutions and employ ML IDS in real-world systems which comprise IoT and cloud networks to enhance performance flexibility and security
PID control based Cluster Autoscaling for Kubernetes
No full textContainerized cloud based microservices are deployed using Kubernetes framework in recent approaches in the industry. Kubernetes offers in-built autoscaling tools like the Horizontal Pod Auto-scaler and Vertical Pod Auto-scaler for provisioning resources based on demand. This research proposes a Control-based autoscaling method implementing the Proportional, Integral and Derivative control theory (PID).the PID control logic is used to address the dynamic resource allocation in an efficient approach. In this research the scaling requirement and decision is aggregated by analysing CPU utilization ,Memory Utilization and Error requests. By passing these metrics into the PID control loop to obtain a stable value for autoscaling the hardware. Kubeadm and terraform are used for deployment and the achieve horizontal and vertical scaling. This research highlights the potential of the PID Auto-scaler to achieve better scaling performance in container based microservices by achieving better cost-effectiveness, response-time and performance. In Future the PID control logic can be implemented on various other metrics together or individually on any number of metrics then calculate an aggregated mean, later the scaling output can be further refined by applying advanced mean models to obtain more stable and effective results