Analysis of Android Applications using Permissions and Intents for Malware Detection

Abstract

With Android’s increasing dominance as the primary mobile operating system, accounting for approximately 85% of the market share, malicious actors have tar geted Android devices more frequently. Traditional antivirus software employs signature-based detection methods that struggle to identify new and evolving malware, necessitating frequent updates to their signature databases. To sur mount this limitation, researchers have proposed Android malware detection models based on machine learning and deep learning. However, previous re search has focused primarily on detecting malware using a singular characteris tic. In this study, I present an innovative method for enhancing Android mal ware detection by combining multiple static features. The methodology involves extracting permission, and intent features from 20,000 APK files. From these ex tracted features, I constructed binary two-dimensional feature vectors, which are then used for training and detection with various machine learning classifiers. In addition, I strived to decrease the time required for training and detection by employing fewer features. I assessed the efficacy of machine learning and deep learning models on both the original and reduced datasets, as well as the im pact on training and detection time. My experimental results demonstrate that Random Forest outperforms other machine learning classifiers, obtaining 96.80% accuracy on the complete set of features and 96.32% accuracy on the reduced set of features for malware classification. RNN obtains the highest accuracy of 96.37 percent on the complete feature set and 95.97 percent on the reduced feature set when employing deep learning models. In addition, the phase of dimensionality reduction, with the help of feature reduction, reduces training and testing time by approximately 46%.