EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Enhancing Malware Family Classification Through Fully Fine‐Tuned Vision Transformers and Attention Map Interpretability

Fahd A. Alqasemi, Mohammed Hameed, Belal Al-Fuhaidi, Malak Korami, Adham Waleed, Abdulahakim-Al Wahabi

Applied Computational Intelligence and Soft Computing · 2026-01 · 2026 (1)

DOI: 10.1155/acis/4731159 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Applied Computational Intelligence and Soft Computing, online erschienen am 16.11.2026. Verfasst von Fahd A. Alqasemi und 5 weiteren Autor:innen. Die Arbeit nennt 19 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Malware has become increasingly complex and threatening over time, posing many risks to individuals and society at large. Hence, it becomes crucial to come up with efficient malware family classification techniques. However, since malware has been evolving in its obfuscation and polymorphism, traditional classification techniques are ineffective. One way to handle the challenges above is by developing an image‐based malware classification method that employs deep learning. Using previous research results obtained from ResNet‐50, this paper evaluates a fully fine‐tuned vision transformer for the purpose of malware family classification based on the RGB images of a PE file. Instead of CNN models used in traditional image‐based malware classification, this study exploits transformers because of their capacity to capture long dependencies as well as salient structural information from malware images through a self‐attention mechanism. This transformer model was trained on the Blended Malware Image dataset that contains 31 malware families with thorough preprocessing and data augmentation steps in order to avoid overfitting. Experimental results from a comprehensive comparative study demonstrate that the proposed ViT‐B/16 model achieves competitive classification accuracy, with a mean of 96.59% ± 0.32% across 3 independent runs with different random seeds, demonstrating high statistical reliability. The model achieved a peak performance of 97.13% in its best‐performing configuration. Compared to state‐of‐the‐art CNN architectures trained under identical conditions—including ResNet‐50 (97.20% ± 0.08%), EfficientNet‐B0 (97.37% ± 0.07%), DenseNet‐121 (97.62% ± 0.23%), and MobileNetV2 (97.18% ± 0.13%)—modern CNNs demonstrate marginally higher accuracy under identical training protocols. However, the vision transformer offers unique advantages through inherent interpretability via attention mechanisms. Attention map visualizations reveal discriminative saliency areas that contribute most to classification decisions, providing security analysts with transparent insights into the model’s decision‐making process. These findings demonstrate that while CNNs achieve competitive accuracy, vision transformers provide valuable complementary capabilities through explainable AI, paving the way toward more transparent and interpretable cybersecurity solutions.

Zitieren

APA

Alqasemi, F. A., Hameed, M., Al-Fuhaidi, B., Korami, M., Waleed, A., & Wahabi, A. (2026). Enhancing Malware Family Classification Through Fully Fine‐Tuned Vision Transformers and Attention Map Interpretability. Applied Computational Intelligence and Soft Computing, 2026(1). https://doi.org/10.1155/acis/4731159

Vancouver

Alqasemi FA, Hameed M, Al-Fuhaidi B, Korami M, Waleed A, Wahabi A. Enhancing Malware Family Classification Through Fully Fine‐Tuned Vision Transformers and Attention Map Interpretability. Applied Computational Intelligence and Soft Computing. 2026;2026(1). doi:10.1155/acis/4731159

BibTeX
@article{Alqasemi2026,
  author = {Alqasemi, Fahd A. and Hameed, Mohammed and Al-Fuhaidi, Belal and Korami, Malak and Waleed, Adham and Wahabi, Abdulahakim-Al},
  title = {Enhancing Malware Family Classification Through Fully Fine‐Tuned Vision Transformers and Attention Map Interpretability},
  journal = {Applied Computational Intelligence and Soft Computing},
  year = {2026},
  volume = {2026},
  number = {1},
  doi = {10.1155/acis/4731159}
}
Kostenlos starten →