LFreeDA Dataset (MB-24+)

Data accompanying "LFreeDA: Label-Free Drift Adaptation for Windows Malware Detection" (ACSAC 2026).

This dataset provides the precomputed features, labels, and intermediate artifacts needed to reproduce the paper's Step I–III pipeline on the MB-24+ malware corpus, spanning five rolling monthly adaptation tasks (July→Aug, Aug→Sep, Sep→Oct, Oct→Nov, Nov→Dec 2024).

Code: https://github.com/gloryer/LFreeDA

Contents

• graph_features/ — control-flow-graph (CFG) embeddings and adjacency matrices per binary, organized by month and malware/benign source.
• image_features/ — CFG-derived image representations per binary, organized by month and malware/benign source.
• labels/ — per-month malware labels (SHA-256 hash, malware family, and binary attack label).
• stepI_trained_models/ — pretrained Step I generator/classifier weights for all five adaptation tasks.
• stepII_constructed_datasets/ — precomputed Step II pseudo-label-filtered training/test sets for all five adaptation tasks, used directly by the Step III Warm-start/AdvDA scripts.

See the code repository's readme.md and artifact/ARTIFACT.md for exact usage instructions and expected results.

License

Released under CC0-1.0 (public domain dedication).

Citation

If you use this dataset, please cite our paper:

@inproceedings{lfreeda2026,
  title     = {LFreeDA: Label-Free Drift Adaptation for Windows Malware Detection},
  author    = {Adrian Shuai Li and Elisa Bertino},
  booktitle = {2026 IEEE Annual Computer Security Applications Conference (ACSAC)},
  year      = {2026}
}