Classification of malicious android applications by machine learning methods using permission properties

This study applies machine learning methods to classify Android applications as malicious or benign using permission features. A dataset consisting of 2,854 malware and 2,870 non-malware apps with 117 features was used. Classification was performed with Adaboost (AB), random forest (RF), and artificial neural networks (ANN), while the information gain (IG) algorithm was used to select relevant features. The classification process was carried out in three steps: first using all 117 features, second with 60 selected features, and third with 20 selected features. The highest accuracy, 98.4, was achieved using 117 features and ANN. The models were evaluated using precision, recall, F1 score, ROC curve, and AUC metrics. Additionally, the training and testing times of all models were analysed. The study also employed correlation and weighted correlation analysis to assess the importance of permission features.