紀要論文

機械学習を用いたAndroidマルウェア検知の再考

RETHINKING OF ANDROID MALWARE DETENCITION USING MACHINE LEARNING

jpn

24368083

https://doi.org/10.15002/00025388

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

6

2022-03-24

Version of Record

The Android OS has been gaining market share in recent years due to the increase in IoT devices and the popularity of smartphones. In general, attackers target systems with a large number of users, and the number of Android malware victims has been increasing. In 2017, it was reported that 750,000 new Android malware were discovered in the first quarter alone. In 2017, 750,000 new Android malware were reported to have been discovered in the first quarter of the year alone, making Android malware an urgent issue to be addressed. In the past, three methods of malware detection were used: surface analysis, dynamic analysis, and static analysis. However, these methods have disadvantages in terms of human and time costs. Recently, in addition to these conventional rule-based methods, detection methods using machine learning have been reported. In particular, methods based on deep learning have been attracting attention. Compared to classical machine learning methods such as logistic regression, methods based on deep learning are said to have higher generalization performance and more accurate detection capability. Many reports on the application of deep learning techniques have shown accuracies of over 90%. However, in the case of Android malware, which is difficult to collect data, it is often evaluated within the same dataset, and overtraining on the dataset is suspected. In this study, we evaluate the generalization performance of the model on both known and unknown datasets. In addition, by comparing classical machine learning and deep learning, we reconsidered the pros and cons of using deep learning. As a result, we found that the discrimination accuracy of the unknown dataset is about 10 15% lower than that of the known dataset. In addition, there was no significant difference in accuracy between classical machine learning and deep learning.

Article