Scholarlink

<p>This study presents an efficient procedure to analyze the efficiency of Machine Learning (ML) models predicting normal or anomalous classes in a military dataset. Ten ML algorithms are categorized into four groups: general, voting, bagging, and boosting classifiers. Furthermore, Without Features Selection (WOF), Correlated Features Selection (COF), and With Features Selection (WF) are three feature selection strategies are utilized. Then, the performance of algorithms is evaluated based on multiple metrics such as Precision, Recall, F1 Score, Accuracy, and Cross-Validation (CV) Score. In addition, statistical tests like Analysis of Variance (ANOVA), Multivariate Analysis of Variance (MANOVA), and Paired Sample T-Test are conducted to gauge the ML algorithms based on single metrics, multiple metrics, and paired measures, respectively. A thorough examination of the military dataset, KDD Cup 1999, is conducted using its two variants, subset (10%) and superset (100%). The results show that feature selection significantly impacts the models’ efficiency. Also, the bagging algorithms, Extra Tree Classifier (ETC), and Random Forest (RF) classifiers perform better than other groups when entangled with the feature selection but not significantly. In fact, none of the groups is dominant over the rest, which indicates a weak correlation of grouping analysis. The best-performing models in each group have more than 90% efficiency for all the metrics for the subset dataset, and they have competent results as compared to the models in the literature. On the other hand, Decision Tree (DT) and RF algorithms fail to impress for the superset evaluation, but K-Nearest Neighbour (KNN) model performs consistently on both the datasets having significant values for every performance metric. It shows the ability of KNN algorithm to rebuild itself with an extended dataset, whereas other promising algorithms need explicit training while the dataset size increases.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Efficient and Robust Procedure of Cyber Attack Detection and Prediction Using Machine Learning Algorithms for Military Data

  • 期刊论文

  • 发表日期：

    2025年8月8日

查看全文
  • Dillip Rout,
  • Vipin Y. Borole,
  • Prasanna Kapse

摘要

This study presents an efficient procedure to analyze the efficiency of Machine Learning (ML) models predicting normal or anomalous classes in a military dataset. Ten ML algorithms are categorized into four groups: general, voting, bagging, and boosting classifiers. Furthermore, Without Features Selection (WOF), Correlated Features Selection (COF), and With Features Selection (WF) are three feature selection strategies are utilized. Then, the performance of algorithms is evaluated based on multiple metrics such as Precision, Recall, F1 Score, Accuracy, and Cross-Validation (CV) Score. In addition, statistical tests like Analysis of Variance (ANOVA), Multivariate Analysis of Variance (MANOVA), and Paired Sample T-Test are conducted to gauge the ML algorithms based on single metrics, multiple metrics, and paired measures, respectively. A thorough examination of the military dataset, KDD Cup 1999, is conducted using its two variants, subset (10%) and superset (100%). The results show that feature selection significantly impacts the models’ efficiency. Also, the bagging algorithms, Extra Tree Classifier (ETC), and Random Forest (RF) classifiers perform better than other groups when entangled with the feature selection but not significantly. In fact, none of the groups is dominant over the rest, which indicates a weak correlation of grouping analysis. The best-performing models in each group have more than 90% efficiency for all the metrics for the subset dataset, and they have competent results as compared to the models in the literature. On the other hand, Decision Tree (DT) and RF algorithms fail to impress for the superset evaluation, but K-Nearest Neighbour (KNN) model performs consistently on both the datasets having significant values for every performance metric. It shows the ability of KNN algorithm to rebuild itself with an extended dataset, whereas other promising algorithms need explicit training while the dataset size increases.