紀要論文

Deep Learning Based Android Malware detection with Android packer classifier

eng

24368083

https://doi.org/10.15002/00025358

法政大学大学院理工学研究科

法政大学大学院紀要. 理工学研究科編

63

1

3

2022-03-24

Version of Record

Deep learning as well as machine learning are grown to be one of mainstream methods of malware detections, including that on Android platforms. However, the impact of packer is often ignored in previous studies. Packers is program that aims at protecting the original code from being analyzed. It could also influence the efficiencies of selected features in machine learning. While some studies noticed this problem, the test and training are made on re-obfuscated samples, which is not convincible. Under this background, we proposed an end to end deep learning based Android malware detection method. The main idea of our method is that let the feature extraction layer to learn features of malware and packers simultaneously from part of the input, which is realized by multi-task learning. And this idea is partly inspired by the fact that unpacking is usually the first step in manual analysis of malware. To avoid re-obfuscation, conventional method is used to get obfuscation labels. The results show that our method averagely achieved 97.0%-99.8% accuracy on different datasets. The experiments also suggest that the use of packer information in multi-tasking learning could help suppress the overfitting.

Article