Answers are grounded in this paper and cite their sources.
Android's permission system evolves as permissions are deprecated, restricted, or reserved for system use, altering the feature space for malware detection and introducing concept drift. We study how these changes affect machine learning models using a dataset of over 70k applications and 166 permissions. Across ML and DL classifiers, removing restricted permissions without considering time reduces accuracy by 1.3–2.1 percentage points, while excluding deprecated or not-for-third-party permissions has marginal effects. When training on one year and testing on others, dataset balancing improves F1 scores and converts many critical-drift cases into regular drift, but also increases the number of drift-affected years across permission groups. Using Kolmogorov–Smirnov tests (F1, p → 0.05), balancing increases significant drift detections on both real devices and emulators. To examine platform evolution, we introduce Android-version–specific detectors and summarize temporal robustness with the AUT and A-AUT metrics, observing lower forward-transfer performance and greater stability for RF and RNNs. Finally, adversarial abuse of deprecated and restricted permissions yields higher attack success under drift, up to 55%, with reduced impact in later years as the permission space stabilizes. Overall, permission evolution, dataset balance, OS versioning, and adversarial manipulation each induce distinct drift patterns, offering quantitative guidance for designing adaptive, permission-aware malware detection models.
No takes yet. Share an insight, caveat, or question.
Sabbah et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: