MVVDroid: Android Malware Detection based on Multi-View Visualization.

With the increasing popularity of the Android platform, more and more malicious apps have been released, threatening the security of mobile users. As they have complex logic and diverse variations, each of the existing schemes can only detect a portion of them. In this paper, we present MVVDroid, an Android malware detection technique based on software visualization and multi-view learning. To capture the semantic information of Android apps comprehensively, we first visualize the apps from three views, including behavior view, operation view and bytecode view. Software visualization helps us to extract the apps’ multi-granularity structural information, which is distinctive for malware detection. We then fuse the comprehensive semantics at different levels to profile the maliciousness of the apps precisely based on multi-view learning. This paper, to the best of our knowledge, is the first work that combines multi-view learning with software visualization to achieve an effective Android malware detection. Experiments show that MVVDroid is more effective than other representative techniques, such as Drebin and a visualization-based method proposed by Sun et al. in 2021, with an accuracy of 97.78% and an F1-score of 97.14%. Furthermore, we compare the effectiveness of different combinations of views to validate the rationality of our view selection.