Android Malware Detection using LSI-based Reduced Opcode Feature Vector
Adarsh Kumar Singh, Gandharv Wadhwa, Mayank Ahuja, Keshav Soni, Kapil Sharma
Abstract
Android's status as the most popular operating system, while catching the attention of legitimate developers, also lures in a mul- titude of malicious content developers making it highly vulnerable to a plethora of malware. Such a scenario has risen mainly as a consequence of the open-source nature of its framework facilitating a significant number of third-party applications to run on Android. To impede any malicious attempts, a number of detection techniques have been proposed covering a wide variety of methodologies. In this paper, we utilize Latent Semantic Indexing, an information retrieval technique, which enables the construc- tion a lower dimension representation of opcodes while preserving the semantic knowledge that they originally encompass, thus allowing to build a lightweight detection system. The effectiveness of this reduced feature set is assessed with the help of multiple machine learning classifiers. In an attempt to augment the performance of the employed classifiers, we additionally incorporate permissions and intents to the feature set. The best results were indeed achieved on this integrated feature set comprising of all the three features with Random Forest which attained an accuracy of 93.92% with 0.9658 ROC-AUC score.