Combat Mobile Malware via N-gram Based Deep Learning

Abstract

Today, mobile devices are beginning to be used in every aspect of life. In addition to being able to perform financial transactions such as banking and shopping, mobile devices can also store personal information such as pictures / videos on these platforms, and important information about the current surroundings of the phone, such as location / sound, can be obtained. Among the mobile platforms, the popularity of the Android operating system and its open source code make it the main target for malware developers. Today's antivirus software is not effective against malicious software that has been tampered with or encountered for the first time while it is effective against pre-existing threats because they are mostly signature-based. New threats need to be detected as quickly as possible, considering new-signatured versions of the same malware can be created quickly and easily with automatic tools. For this reason, researches based on machine learning and deep learning have been conducted in the last years. In this study deep learning methods, which have been tried to be used successfully in all areas of life in recent years, are tested in mobile malware detection. The opcodes of Android applications were grouped in groups of 2 and 3, their features were extracted, weights were optimized using stacked denoising auto encoder and classified by a multi-layered artifical neural network. As a result of the classification, harmful software was detected with an accuracy of 92.04%