Predicting malicious activity in Android using Machine Learning

The extent to which malware camouflages has begged the attention of a wide range of researchers all over the world to direct their attention to developing effective detection methods for the same. Current antiviral techniques, which are hash-based and heuristic-driven, do not appear to handle the problem efficiently or effectively. The remarkable success of machine learning in tackling this challenge has increased its usage in developing these detection systems. In this paper, we examine such approaches in order to assess the accuracy of ML-based methods, particularly in a supervised learning scenario, and to provide the reader with a systematization of knowledge on the current state of the field. In this paper, three different methods were proposed based on machine learning models to detect malicious programs, and we evaluated them to show their efficacy and efficiency.

I. Introduction

To be able to detect effectively malware programs through Supervised Machine Learning approaches, we needed some really big labeled datasets to analyze. We started from the DREBIN dataset used in the DREBIN paper [1], a 123K labeled Android application dataset, containing 5,560 malwares, resulting in a rate of 95.4% benign programs.