Multi-Line Defense Against Windows Adversarial Malware by Using Windows PE Information
IEEE Access · 2026 · 14 · 95716-95730
DOI: 10.1109/access.2026.3705939 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Hannah Ho und 2 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Research Foundation of Korea.
Abstract (Originalsprache)
Deep learning has recently been in the spotlight among malware detection researchers in the sense that its training-based robust decision process can lead to efficient and effective malware detection. However, this training-based nature can also provide an attacker with a chance of being able to evade the detection. More specifically, an attacker can keep perturbing malware and testing the perturbed malware against the pre-trained malware detection system until the perturbed malware successfully bypasses it. We call this perturbed malware adversarial malware. A variety of adversarial malware has been recently proposed in the literature. When compared to many different types of adversarial malware, defensive approaches against this adversarial malware have been relatively underexplored in the literature. The most representative defense is to retrain the pre-trained malware detection system with diverse types of the adversarial malware or train it with ablated versions of original malware. However, these approaches need to perform retraining or training task whenever new types of adversarial malware are gathered and analyzed. To pacify the limitations of these approaches, we propose a non-training-based multi-line defense scheme against Windows adversarial malware that leverages Windows PE information. The main rationale behind focusing on Windows adversarial malware detection is that diverse types of adversarial malware originate mainly from Windows original malware. The multi-line defense starts with the first defense of header information manipulation. In case of failure of the first defense, it continues to the second defense of byte-block-check-based ablation of null padding regions. The final defense, selective ablation of sections, is performed in case of the collapse of the second defense line. This multi-line defense is employed to detect adversarial malware attacks in which perturbation code is embedded in header, null padding regions, appended sections of Windows PE files. We implement our multi-line defense scheme in Python, evaluate its detection performance against diverse types of adversarial malware, and perform comparison with the state-of-the-art (de) randomized smoothing schemes. Our evaluation results demonstrate that our multi-line defense achieves a detection rate ranging from 95.1% to 100% while maintaining a false positive error rate ranging from 0.51% to 8.9%. Note that the high detection rates against adversarial ransomware are achieved with the relatively small sample size ranging from 61 to 224 when compared to other types of adversarial malware. Moreover, our multi-line defense outperforms (de) randomized smoothing schemes in terms of the detection rate against all types of adversarial malware and the false positive rate for all types of benign PE files with the exception of one adversarial malware type.
Förderung
- National Research Foundation of Korea
Zitieren
APA
Ho, H., Ho, J., & Ho, S. (2026). Multi-Line Defense Against Windows Adversarial Malware by Using Windows PE Information. IEEE Access, 14, 95716-95730. https://doi.org/10.1109/access.2026.3705939
Vancouver
Ho H, Ho J, Ho S. Multi-Line Defense Against Windows Adversarial Malware by Using Windows PE Information. IEEE Access. 2026;14:95716-95730. doi:10.1109/access.2026.3705939
BibTeX
@article{Ho2026,
author = {Ho, Hannah and Ho, Jun-Won and Ho, Sungjin},
title = {Multi-Line Defense Against Windows Adversarial Malware by Using Windows PE Information},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {95716--95730},
doi = {10.1109/access.2026.3705939}
}
EUVIMED