Concept drift and cross-device behavior: Challenges and implications for effective android malware detection

The large body of Android malware research has demonstrated that machine learning methods can provide high performance for detecting Android malware. However, the vast majority of studies underestimate the evolving nature of the threat landscape, which requires the creation of a model life-cycle to ensure effective continuous detection in real-world settings over time. In this study, we modeled the concept drift issue of Android malware detection, encompassing the years between 2011 and 2018, using dynamic feature sets (i.e., system calls) derived from Android apps. The relevant studies in the literature have not focused on the timestamp selection approach and its critical impact on effective drift modeling. We evaluated and compared distinct timestamp alternatives. Our experimental results show that a widely used timestamp in the literature yields poor results over time and that enhanced concept drift handling is achieved when an app internal timestamp was used. Additionally, this study sheds light on the usage of distinct data sources and their impact on concept drift modeling. We identified that dynamic features obtained for individual apps from different data sources (i.e., emulator and real device) show significant differences that can distort the modeling results. Therefore, the data sources should be considered and their fusion preferably avoided while creating the training and testing data sets. Our analysis is supported using a global interpretation method to comprehend and characterize the evolution of Android apps throughout the years from a data source-related perspective.