HiFAB Preview

Dataset Summary

This repository is a Hugging Face release draft for HiFAB Preview. The current package is centered around protocol manifests and image-only split files for binary face anti-spoofing / presentation attack detection.

• Total indexed records across all manifests: 110538
• Modality: image
• Protocol families: p1, p2, p3
• Splits per config: train, validation (from dev), test
• Binary labels: 0 = live, 1 = attack

The release currently assumes that the raw assets are distributed alongside these files, for example as Data_preview.zip or as an extracted Data_preview/ directory in the final dataset repository.

Repository Layout

Recommended repository contents:

HiFAB/
|-- README.md
|-- data/
|   |-- p1_image/
|   |   |-- train-00000-of-00001.parquet
|   |   |-- validation-00000-of-00001.parquet
|   |   `-- test-00000-of-00001.parquet
|   |-- p2_image/
|   |   |-- train-00000-of-00001.parquet
|   |   |-- validation-00000-of-00001.parquet
|   |   `-- test-00000-of-00001.parquet
|   `-- p3_image/
|       |-- train-00000-of-00001.parquet
|       |-- validation-00000-of-00001.parquet
|       `-- test-00000-of-00001.parquet
|-- Data_preview/
|   `-- ...
`-- protocol/
    `-- *.txt

Configs And Splits

Each Hugging Face config corresponds to one protocol in the image-only release:

Config	Split	Records	Live (0)	Attack (1)
p1_image	validation	7632	197	7435
p1_image	train	14964	383	14581
p1_image	test	14250	367	13883
p2_image	validation	5088	13	5075
p2_image	train	20614	839	19775
p2_image	test	11144	95	11049
p3_image	validation	5396	13	5383
p3_image	train	22367	839	21528
p3_image	test	9083	95	8988

Manifest Fields

Each Parquet record contains the following columns:

• relative_path: sample path relative to the dataset root.
• label: binary class label.
• label_name: human-readable label (live or attack).
• protocol: evaluation protocol family.
• modality: image.
• split: original split name (train, dev, or test).
• coarse_group: high-level source branch extracted from the path (for example Live, Physical, Digital).
• level_2, level_3, level_4: additional path-derived grouping levels retained for filtering and auditing.
• file_name: terminal file or sample identifier.

Composition Snapshot

Top-level path groups observed in the current preview:

• Digital: 99840 indexed records
• Physical: 7857 indexed records
• Live: 2841 indexed records

Suggested Citation

@dataset{hifab_preview_2026,
  title = {HiFAB Preview},
  author = {HiFAB Team},
  year = {2026},
  note = {Preview release}
}

Licensing And Access

Replace this section before publication:

• Add the exact license for the released preview assets and manifests.
• Confirm whether the preview can be redistributed publicly on Hugging Face.
• If any component originates from third-party benchmarks or prior datasets, cite them explicitly and state the redistribution basis.

Responsible AI Considerations

This repository includes a generated croissant_rai_overlay.jsonld starter file, but it still requires human review. At minimum, fill in:

• how the live and attack data were collected,
• whether any human subjects are represented,
• whether identities, demographics, regions, or sensors are unevenly distributed,
• the intended evaluation use cases,
• known risks, failure modes, and non-recommended uses,
• the plan for updates, corrections, or takedown requests.