Deep Learning-Based Malware Detection Under Real-World Constraints: A Systematic Review of Class Imbalance, Concept Drift, and Interpretability
Faculty of Natural and Applied Sciences Journal of Computing and Applications · 31.03.2026 · 3 (1) · 87-92
DOI: 10.63561/jca.v3i1.1206 ↗Kurz gefasst
Systematische Übersicht in der Zeitschrift Faculty of Natural and Applied Sciences Journal of Computing and Applications, online erschienen am 31.03.2026. Verfasst von Abubakar Bodinga Bodinga und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Malware detection remains one of the most persistent and complex challenges in cybersecurity. The rapid evolution of attack techniques, fueled by the professionalization of cybercrime, continually outpaces traditional defenses. While deep learning has significantly enhanced detection capabilities, its real-world deployment is critically hampered by three interconnected and often overlooked challenges: extreme class imbalance, concept drift (including adversarial evolution), and the interpretability gap of black-box models. This Systematic Literature Review (SLR) synthesizes state-of-the-art research from 2014 to 2025 on malware detection across static, dynamic, and hybrid analysis methods, with a focused analysis on these three constraints. Following a PRISMA-guided methodology, this review analyzes 162 high-quality studies. It reveals that while research has progressed from foundational deep learning applications to advanced solutions like generative augmentation for imbalance, self-supervised test-time adaptation for drift, and integrated explainable AI (XAI) pipelines critical gaps persist. Our synthesis yields five key insights: (1) deep learning enhances accuracy but remains brittle under real-world data imbalance and adversarial drift; (2) current drift adaptation strategies, including recent federated and hybrid approaches, seldom holistically model adversarial intent; (3) GAN-based augmentation improves minority-class detection but lacks robust, security-focused evaluation of synthetic samples; (4) interpretability studies, despite recent integration efforts, remain fragmented and are rarely validated with human analysts to ensure actionable intelligence; and most critically, (5) no existing architecture jointly and seamlessly integrates continuous drift adaptation, dynamic imbalance correction, and operational interpretability. This review not only maps the evolution of these challenges but also crystallizes the pressing need for a unified framework. It provides the foundational justification for the proposed MAD-FIT (Malware Adaptive Detection with Fusion, Interpretation, and Training Dynamics) framework, which is designed to bridge these gaps and advance the field toward robust, adaptive, and trustworthy next-generation malware detection systems.
Zitieren
APA
Bodinga, A. B., Garko, A. B., Ibrahim, N. M., & Gabi, D. (2026). Deep Learning-Based Malware Detection Under Real-World Constraints: A Systematic Review of Class Imbalance, Concept Drift, and Interpretability. Faculty of Natural and Applied Sciences Journal of Computing and Applications, 3(1), 87-92. https://doi.org/10.63561/jca.v3i1.1206
Vancouver
Bodinga AB, Garko AB, Ibrahim NM, Gabi D. Deep Learning-Based Malware Detection Under Real-World Constraints: A Systematic Review of Class Imbalance, Concept Drift, and Interpretability. Faculty of Natural and Applied Sciences Journal of Computing and Applications. 2026;3(1):87-92. doi:10.63561/jca.v3i1.1206
BibTeX
@article{Bodinga2026,
author = {Bodinga, Abubakar Bodinga and Garko, Ahmed Baita and Ibrahim, Nuruddeen Mahmud and Gabi, Danlami},
title = {Deep Learning-Based Malware Detection Under Real-World Constraints: A Systematic Review of Class Imbalance, Concept Drift, and Interpretability},
journal = {Faculty of Natural and Applied Sciences Journal of Computing and Applications},
year = {2026},
volume = {3},
number = {1},
pages = {87--92},
doi = {10.63561/jca.v3i1.1206}
}
EUVIMED