Policing Android Malware using Object-Oriented Metrics and Machine Learning Techniques

Abstract

The primary motive of a malware is to compromise and exfiltrate sensitive user data from a system generally designed to uphold the fundamental principles of information security i.e., confidentiality, integrity and availability. Android being the most widely used mobile operating system, is a lucrative ground for malware designers in leveraging system flaws to gain unauthorized user information access. In order to attenuate these issues, it is imperative to design and build robust automated tools for effective android malware prediction. In this paper we bring forward a novel method for android malware detection using object-oriented software metrics and machine learning techniques. 5,774 android apps are collected from Androzoo repository, then it’s software metrics are extracted and aggregated using sixteen aggregationmeasures which forms the basis of our metrics-based dataset. A total of three hundred and four different machine-learned models are built using various data-sampling techniques, feature-selection methods and machine learning algorithms. Finally, a machine learned model built using SVMSMOTE data-sampling technique applying SPM (Significant Predictor Metrics) feature selection methods over GDCG2H (Conjugate Gradient with Powell/Beale Restarts and 2 Hidden Layers) machine learning algorithm, yields a better malware predictor with AUC (area under ROC curve) value of 0.86.