Scholarlink

Malware has become a massive problem for Android due to its extensive market share with traditional signature-based detection methods unable to keep up with new and varied threats. In this paper, the authors propose a machine-learning framework for Android malware detection confusion on a hybrid set of static and dynamic features extracted from 50 000 Android apps. This preprocessing limited the dataset to 30000 samples and the selection of features was initially based on correlation maximum-rank descriptor and heatmap analysis of the features. The classifiers that were used are an optimized version of Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest and the bootstrap aggregating meta-classifier Gradient Boosting which boasts over 98% accuracy on the processed dataset In-depth analysis and extensive experiments were carried out to compare between the different stacking systems where it proved that the stacking methodology takes better performance over large-scale real-word datasets. Future work will look into the autoencoder and generating neural networks for the second-stage detection of unseen classes of malware and the applicability of the system in other datasets.

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

Comprehensive Malware Detection Framework Using Stacked Machine Learning Classifiers

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Nuha Husham Alhadethy,
  • Teba Mustafa Bahya,
  • Asfan Muqtadir

摘要

Malware has become a massive problem for Android due to its extensive market share with traditional signature-based detection methods unable to keep up with new and varied threats. In this paper, the authors propose a machine-learning framework for Android malware detection confusion on a hybrid set of static and dynamic features extracted from 50 000 Android apps. This preprocessing limited the dataset to 30000 samples and the selection of features was initially based on correlation maximum-rank descriptor and heatmap analysis of the features. The classifiers that were used are an optimized version of Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest and the bootstrap aggregating meta-classifier Gradient Boosting which boasts over 98% accuracy on the processed dataset In-depth analysis and extensive experiments were carried out to compare between the different stacking systems where it proved that the stacking methodology takes better performance over large-scale real-word datasets. Future work will look into the autoencoder and generating neural networks for the second-stage detection of unseen classes of malware and the applicability of the system in other datasets.