This paper provides an in-depth analysis of Android malware that bypassed the strictest defenses of the Google Play application store and penetrated the official Android market between January 2016 and July 2021. We systematically identified 1,238 such malicious applications, grouped them into 134 families, and manually analyzed one application from 105 distinct families. During our manual analysis, we identified malicious payloads the applications execute, conditions guarding execution of the payloads, hiding techniques applications employ to evade detection by the user, and other implementation-level properties relevant for automated malware detection. As most applications in our dataset contain multiple payloads, each triggered via its own complex activation logic, we also contribute a graph-based representation showing activation paths for all application payloads in form of a control- and data-flow graph. Furthermore, we discuss the capabilities of existing malware detection tools, put them in context of the properties observed in the analyzed malware, and identify gaps and future research directions. We believe that our detailed analysis of the recent, evasive malware will be of interest to researchers and practitioners and will help further improve malware detection tools.

This paper investigates the problem of classifying Android applications into malicious and benign. We analyze the performance of a popular malware detection tool, Drebin, and show that its correct classification decisions often stem from using benign rather than malicious features for making predictions. That, effectively, turns the classifier into a benign app detector rather than a malware detector. While such behavior allows the classifier to achieve a high detection accuracy, it also makes it vulnerable to attacks, e.g., by a malicious app pretending to be benign by using features similar to those of benign apps. In this paper, we propose an approach for deprioritizing benign features in malware detection, focusing the detection on truly malicious portions of the apps. We show that our proposed approach makes a classifier more resilient to attacks while still allowing it to maintain a high detection accuracy.