Atomic-Inspired Hybrid Feature Model for Robust Android Malware Detection

Abstract We created a hybrid feature framework for finding Android malware that combines static and dynamic analysis methods. Static features are taken from APK metadata and DEX bytecode images, while dynamic behaviors are recorded using sequences of runtime API calls. The combination of the different features has given us a better picture of how Android apps work. We have added an atomic-inspired design of features to the framework for modeling both the structural and behavioral traits of applications. This is helping to find malware more easily. All application samples come from the AndroZoo repository, which is open to the public. A machine learning pipeline is used, and it is combined with image-based CNN embeddings structured features to use an XGBoost classifier to tell the difference between good and bad apps. Our experimental findings indicate a precision of 99.93\%, highlighting the resilience of the proposed hybrid detection methodology.