Work

Ensemble Learning for Android Malware Detection: Stability Across Malware Categories and Feature Representations

Year: 2026
Type: article
Abstract: Android malware categories exhibit distinct behaviors, yet existing studies typically evaluate one ensemble strategy or feature representation, providing limited evidence of cross-category stability. This study compares a standalone decision tree, six homogeneous ensembles, and three heterogeneous ensembles across Adware, Banking malware, SMS malware, Riskware, and Combined Malware detection using static, dynamic, and hybrid features from CICMalDroid2020. With hybrid features, Stacking achieved mean Matthews correlation coefficient (MCC) values of 0.971–0.996 across the five tasks. Stacking and Soft Voting had the lowest descriptive average ranks, at 2.7 and 2.9, while several boosting ensembles achieved competitive MCC values. Adware and SMS malware were easier to detect than Banking malware and Riskware. Category-specific detectors achieved higher mean MCC in all 40 ratio-matched comparisons, with greater target-category training exposure. Hybrid features had the lowest descriptive average rank across the 50 model–task settings (1.28), followed by static (1.92) and dynamic (2.80) features. Their observed benefit over static features was strongest for Banking and Combined Malware. For hybrid features, Stacking’s median fitting time was about 5.5 times that of voting, while batch inference medians were close. The results describe observed ensemble performance and consistency, category-specific training outcomes, feature-representation differences, and computational trade-offs. These findings are limited to CICMalDroid2020 under a fixed 2:1 benign-to-malware ratio, and generalizability to other datasets and deployment conditions remains to be established.
Language: English
Subfield: Software
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7226022947
FWCI:
Cites: 24
Cited by:
Related to: 10