Machine Learning-Based Android Malware Detection

The use of mobile phones, particularly smartphones, has been growing exponentially in recent times. From 2016 to 2021, smartphone users increased by more than 70%. With the increase in the popularity of smartphones, smartphones have become the prime target for criminal hackers. As a result, Android malware samples are coming to the market at an alarming rate. A study shows that there are more than 4 million malicious Android apps in the market, and each day around 11,000 new malwares add to this number. To combat this mass number of malware, we need a malware detection system that is efficient in detecting malicious Android apps. There are numerous existing malware detection systems, but most of them require countless features from both dynamic and static analysis. Thus, they are not scalable, lightweight, and efficient in detecting malware. Additionally, most studies that used limited features like only permission data, had done their research on much older dataset. Hence, there is a need for new research on this topic. In this paper, we build a permission-based malware detector for Android application with a new dataset and significantly less permissions. Initially, we used support vector machine (SVM) and all the extracted permission data as features to build our classification model. The model accuracy, precision, recall and F1 score were 97.41 percent, which is higher than the other state of the art similar approaches done on an older dataset. Next, we replicated this similar study with a few different machine learning algorithms: random forest, decision tree and logistic regression, and observed they all give similar results. However, tree-based algorithm performs a little better than the other algorithms. Finally, to achieve a lightweight malware detection system, we reduced the number of permissions or the features on a two-step process, and found only a slight difference in results. In the first step, even after reducing the number of permissions by about 94%, the accuracy dropped by only 2.7%. In the second step, we further reduced the number of features or permissions and observed the difference in results. We managed to prune to 9 permissions while maintaining accuracy of 93%, which is lower than technique mentioned in other literature to reduce features.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.