Image-Based Android Malware Detection Using Deep Learning

Abstract

The Android operating system (OS) dominates the mobile phone OS industry, with over 70% of the market share. With the growth of Android OS-based smartphones, it has become a prime target for mobile malware attacks. Minimal alterations in malware samples can easily evade traditional detection methods such as signature-based detection. In contrast, artificial intelligence (AI) and machine learning (ML)-based malware detection has proven more effective, as it can detect zero-day malware. Previous studies have shown that AI/ML-based malware classifiers trained on categorical features are vulnerable to adversarial evasion attacks. Therefore, in this study, we transform the features extracted from Android apps into image-based data and investigate the performance of Convolutional Neural Networks (CNNs), Inception Networks, and Residual Networks (ResNet) on this data. We employ 41,382 Android malware samples belonging to 240 malware families and 36,755 benign apps to train and test the models. Our experiment results show that CNNs outperform Inception Networks and ResNet with up to 99% classification accuracy. Furthermore, our analysis also indicates that CNNs trained on image-based Android malware and benign data outperform various Android malware detection techniques proposed in the literature.