Rapid degradation of linear Android Malware detection under white-box feature-space perturbations

Abstract

This paper presents the comprehensive head-to-head adversarial robustness evaluation of Elastic-Net, Huber-loss logistic regression, and their simple ensemble on static features from ~100,000 imbalanced AndroMD Android apps. White-box ℓ∞ PGD attacks at ε=0.10 cause 68–84% F1-score collapse, with recall most severely degraded. Huber loss provides modest gains (+4–9% F1 over Elastic-Net at moderate ε), while the ensemble offers only marginal improvement. Iterative PGD consistently outperforms FGSM. Full attack grid results are publicly released as a challenging, reproducible baseline for future defenses in lightweight Android malware detection.