Multimodal Deep Learning for Android Malware Classification

/src/ contains the source code for this project. The assemble_dataset.py script assembles the multimodal dataset consisting of MalNet images and graphs. The Jupyter notebook experiments.ipynb includes the experiments and results presented in our paper:

• CNNs on binary images
  • Semantic information encoding (RGB vs greyscale)
  • Transfer learning (ImageNet)
  • Plain CNN grid optimisation
  • Advanced CNN architectures (ResNet18, DenseNet121, MobileNetV2)
• GNNs on function call graphs (GCN, GIN)
• Fusion strategies (2 intermediate, 2 late)
• Optimising late-fusion ensembles (meta-classifier)
  • Grid optimisation
    • Plain CNN + GCN
    • Plain CNN + GIN
  • Bayesian optimisation
    • MobileNetV2 + GCN
    • DenseNet121 + GCN
    • DenseNet121 + GIN
    • MobileNetV2 + DenseNet121 + GCN + GIN
• Tables and figures
  • Fusion strategies
  • Base and late-fusion models
  • Confusion matricies (plain CNN + GCN, DenseNet121 + GIN)
  • UMAP anlaysis (DenseNet121 + GIN)
  • ROC curve (DenseNet121 + GIN)
  • SHAP graphs (DenseNet121 + GIN)

Code authored by James Arrowsmith — please direct correspondence to [email protected]

Citation

Arrowsmith, J.; Susnjak, T.; Jang-Jaccard, J. Multimodal Deep Learning for Android Malware Classification. Mach. Learn. Knowl. Extr. 2025, 7, 23. https://doi.org/10.3390/make7010023

@article{arrowsmith2025multimodal,
  author    = {Arrowsmith, James and Susnjak, Teo and Jang-Jaccard, Julian},
  title     = {Multimodal Deep Learning for Android Malware Classification},
  journal   = {Machine Learning and Knowledge Extraction},
  year      = {2025},
  volume    = {7},
  number    = {1},
  pages     = {23},
  doi       = {10.3390/make7010023},
  url       = {https://doi.org/10.3390/make7010023}
}