Lwin Khin Shar, Biniam Fisseha Demissie, Mariano Ceccato, Yan Naing Tun, David Lo, Lingxiao Jiang, Christoph Bienert

Experimental comparison of features, analyses, and classifiers for Android malware detection


Android malware detection has been an active area of research. In the past decade, several machine learning-based approaches based on different types of features that may characterize Android malware behaviors have been proposed. The usually-analyzed features include API usages and sequences at various abstraction levels (e.g., class and package), extracted using static or dynamic analysis. Additionally, features that characterize permission uses, native API calls and reflection have also been analyzed. Initial works used conventional classifiers such as Random Forest to learn on those features. In recent years, deep learning-based classifiers such as Recurrent Neural Network have been explored. Considering various types of features, analyses, and classifiers proposed in literature, there is a need of comprehensive evaluation on performances of current state-of-the-art Android malware classification based on a common benchmark.

In this study, we evaluate the performance of different types of features and the performance between a conventional classifier, Random Forest (RF) and a deep learning classifier, Recurrent Neural Network (RNN). To avoid temporal and spatial biases, we evaluate the performances in a time- and space-aware setting in which classifiers are trained with older apps and tested on newer apps, and the distribution of test samples is representative of in-the-wild malware-to-benign ratio. Features are extracted from a common benchmark of 7,860 benign samples and 5,912 malware, whose release years span from 2010 to 2020. Among other findings, our study shows that permission use features perform the best among the features we investigated; package-level features generally perform better than class-level features; static features generally perform better than dynamic features; and RNN classifier performs better than RF classifier when trained on sequence-type features.

Link to the paper.