Image-based Android Malware Detection Models using Static and Dynamic Features

Today smartphones are an indispensable part of our everyday activities and store a plethora of sensitive as well as personal information. However, this information is an attractive target of malware designers that can be validated with an ever-increasing number of smartphone malware. Recently researchers explored deep learning for detecting android malware and have seen encouraging results. In this paper, we propose an effective image-based android malware detection system. We used both static and dynamic analysis of android applications to extract six different features: intent, opcode, permission from static analysis, and unigram, bigram, trigram from system call log using dynamic analysis. Then, we proposed a custom malware detection model (MalCNN) that uses static features and achieved accuracy and AUC of 99.56% and 0.99 respectively in malware detection. We also explored MobileNetV2 based malware detection models for dynamic features that achieved accuracy and AUC of 99.85% and 0.99 respectively in malware detection. Our experimental results show that image representation of static or dynamic features can be used for effective malware detection.