Authors:
A. Talbi
1
;
2
;
A. Viens
3
;
L.-C. Leroux
3
;
M. François
3
;
M. Caillol
3
and
N. Nguyen
1
;
4
Affiliations:
1
ETIS Laboratory, CY Cergy Paris University, Cergy, France
;
2
Pôle Judiciaire de la Gendarmerie Nationale, Cergy, France
;
3
CY Tech, CY Cergy Paris University, Cergy, France
;
4
Léonard de Vinci Pôle Universitaire, Research Center, Paris La Défense, France
Keyword(s):
Android Malware Detection, Static Analysis, Feature Importance, URL Embedding, Deep Neural Network.
Abstract:
Effective and efficient malware detection is key in today’s world to prevent systems from being compromised, to protect personal user data, and to tackle other security issues. In this paper, we worked on Android malware detection by using static analysis features and deep learning methods to separate benign applications from malicious ones. Custom feature vectors are extracted from the Drebin and the AndroZoo dataset and different data science methods of feature importance are used to improve the results of Deep Neural Network classification. Experimental results on the Drebin dataset were significant with 99.31% accuracy in malware detection. We extended our work on more recent applications with a complete pipeline for the AndroZoo dataset, with about 40,000 APKs used from 2014 to 2021 pre-tagged as reported malicious or not. The pipeline includes static features extracted from the manifest file and bytecode such as suspicious behaviors, restricted and suspicious API calls, etc. Th
e accuracy result for AndroZoo is 97.7%, confirming the power of deep learning on Android malware detection.
(More)