A Multimodal Deep Network Model for Android Malware Detection Using Permission

The Android platform has grown rapidly over the recent years. Nevertheless, the variant malicious attacks are also increasing due to its popularity and flexibility. Common malicious behaviors, such as privacy and sensitive information theft, pose a serious threat to the economic security and privacy security of users. Android uses a permission-related access control mechanism to limit the operations that a process can perform. In this paper, we propose a multimodal malware detection model MDNMDroid to mine the potential relationship between permissions by combining two different networks. Compared with single network, multimodal network can have more powerful learning ability and filter out more meaningful features for distinguishing malicious and benign samples. Evaluation results based on collected permission dataset demonstrate that MDNMDroid achieves 93.18% accuracy. Besides, we compare our malware detection model with the state-of-the-art related approaches, containing the popular deep learning models and the classical machine learning models. The comparative experimental results further show that MDNMDroid is an effective method in malware detection task.