Multimodal Feature Selection for Android Malware Detection Classifiers

Android applications are fast evolving across the mobile ecosystem, yet Android malware is always emerging. Many researchers have looked into the issue of Android malware detection and proposed hypotheses and methods from various angles. Machine learning appears to be an efficient and promising technique to detect Android malware, according to existing studies. Our paper proposes a simple yet feasible approach to detect android malware using existing machine learning methodologies. We utilize a novel approach for considering multiple appropriate features called the Multivariate Feature Ranking algorithm that focus on deriving the most suitable features which is then applied to classifier models viz Decision Tree, Ada Boost ensemble method, Random Forest and Multi-layer perceptron. We were able to achieve an accuracy of 92% at a high level for the multi-layer perceptron model considering the Drebin dataset under experiment. Putting forth machine learning as the key aspect, we performed data acquisition, data pre-processing, novel feature selection methodology, followed by multiple classifier models that were evaluated and analyzed for further interpretation. The experimental results revealed that the modified feature selection method could help improve classification accuracy for baseline models, which indeed leads to future scope of improvement with advanced deep learning feature learning techniques.