Summary
IEICE Information and Communication Technology Forum
2018
Session Number:SESSION01
Session:
Number:SESSION01_3
Feature Selection Scheme for Android ICC-related Features Based on the Gap of the Appearance Ratio
Kyohei Osuge, Hiroya Kato, Shuichiro Haruta, Iwao Sasase,
pp.-
Publication Date:2018/8/31
Online ISSN:2188-5079
DOI:10.34385/proc.32.SESSION01_3
Summary:
Android malwares are rapidly becoming a potential threat to users. Among several Android malware detection schemes, the scheme using Inter-Component Communication (ICC) is gathering attention. That scheme extracts numerous ICC-related features to detect malwares by machine learning. In order to mitigate the degradation of detection performance caused by redundant features, Correlation-based Feature Selec- tion (CFS) is applied to feature before machine learning. CFS selects useful features for detection in accordance with the theory that a good feature subset has little correlation with mutual features. However, CFS may remove useful ICC-related features because of strong correlation between them. In this paper, we propose a feature selection scheme for Android ICC-related features based on the gap of the appearance ratio. We argue that the features frequently appearing in either benign apps or malwares are useful for malware detection, even if they are strongly correlated with each other. To select useful features based on our argument, we introduce the proportion of the appearance ratio of a feature between benign apps and malwares. Since the proportion can represent whether a feature frequently appears in either benign apps or malwares, this metric is useful for feature selection based on our argument. Unfortunately, the proportion is ineffective when a feature appears only once in all apps. Thus, we also introduce the difference of the appearance ratio of a feature between benign apps and malwares. Since the difference simply represents the gap of the appearance ratio, we can select useful features by using this metric when such a situation occurs. By computer simulation with real dataset, we demonstrate our scheme improves detection accuracy by selecting the useful features discarded in the previous scheme.