Strengthening Android Malware Detection: from Machine Learning to Deep Learning

Abstract

In the recent era of modern world, Android malware continues to escalate, the challenges associated with its usage are growing at an unprecedented rate. This cause a rapid growth in Android malware infections points to an alarming and swift rise in their prevalence, signalling a cause for concern. Traditional anti-malware systems, reliant on signature-based detection, prove inadequate in addressing the expanding scope of newly developed malware. Various strategies have been introduced to counter the escalating threat in the Android mobile field, with many leaning towards machine learning (ML) models limited by a constrained set of features. This paper introduces a novel approach employing a deep learning (DL) framework, incorporating a significant number of diverse features. The proposed framework uses Deep Neural Network (DNN) techniques on OmniDroid dataset, comprising 25,999 features extracted from 22,000 Android Package Kits (APKs). Of these, 16,380 features are meticulously selected for analysis, encompassing Permission, Opcodes, API calls, System Commands, Activities, and Services. Additionally the data is partitioned feature wise and subjected to feature selection on each feature set to ensure equitable consideration of all features. A comparative analysis is presented by comparing the framework accuracy with the accuracies produced by the existing ML models. The presented framework demonstrates notable enhancements in detection accuracy, achieving 89.04% accuracy, attributed to the incorporation of a substantial number of features.