Suchul Lee, Journal of Internet Computing and Services, Vol. 26, No. 6, pp. 33-42, Dec. 2025
10.7472/jksii.2025.26.6.33, Full Text:  HTML
Keywords: Semi-Supervised Learning (SSL), Android malware detection, Consistency Training, Segmentation-Based Network, Label-scarcity Problem

Abstract

Cybersecurity faces escalating label scarcity due to rapidly changing threats and the continual emergence of new malware. To address this setting, we propose a semi-supervised learning (SSL) method for Android malware detection. The approach re-purposes the semantic-segmentation model DeepLabV3+ to extract pixel-level, multi-scale features from image-formatted APK files, and improves detection by enforcing prediction consistency between two parallel branches that receive weakly augmented views of the same input. On the CICMalDroid2020 dataset, across a range of label budgets (Ratio of Labeled, ROL), the proposed method shows consistent superiority over a supervised model and a recursive pseudo-labeling baseline and exhibits stable convergence. An analysis of learning curves indicates that the method effectively leverages structural signals from unlabeled data, enhancing training stability and generalization.

Statistics
Cite this article