Tanko Yahaya Mohammed , Abdulrazaq Abdulrahim , Muhammad Aminu Umar , Aliyu Muhammad Kufena , and Hadiza Isa Abdullahi

Keywords: Empirical Analysis, Accuracy Prediction, Features Selected, Training Data, ML models, Cloud Security

Abstract

Background: Machine learning (ML) techniques have proven to be very effective in providing security in a cloud environment considering the continuous evolving nature of threats. Some of the factors that influence the accuracies of ML models include the specific ML algorithm used, sample size, the number of features selected and portion of dataset used for training. Many studies have conducted empirical analyses of the effects of one or more combination of these factors on predicted accuracies of ML models. However, the effect of the portion of the entire dataset that is used for training the ML model as well as the number of features extracted from the dataset in predicting the accuracy of an ML model is yet to be investigated. Aim This study uses Ordinary Least Square (OLS) regression to investigate if the number of features selected and the size of training data are useful in predicting the accuracies obtained in ML based approaches to cloud security. Method: For this research, we have two independent variables (number of features selected and the size of training data) and one dependent variable (accuracy). We initially selected 16 (sixteen) studies conducted within the last 5 (five) years for our study. We extracted the number of features used, the size of the training data and the accuracies obtained from these studies. After identifying and discarding outliers from the extracted values, we were left with 12 (twelve) studies. We conducted our analysis on these 12 studies. Results: The result of our analysis shows that there exist a weak positive and negative relationships among the dependent and independent variables. Although, our analysis shows weak positive and negative relationships among the variables, our model is useful in predicting the accuracies of ML models given the number of features selected and the size of the training data