imDalton/cutclean-datasets

CutClean: balanced datasets

Data accompanying:

Leonardo Magliolo, Vito Paolo Pastore, Giuseppe Valenzise, Enzo Tartaglione. CutClean: Neural Network Pruning for Privacy-Preserving Inference. Pattern Recognition — Proceedings of the 28th International Conference on Pattern Recognition (ICPR 2026), Lyon, France. Lecture Notes in Computer Science, Springer Nature Switzerland, pp. 450–465. doi:10.1007/978-3-032-31452-9_30

Code: https://github.com/MaglioloLeonardo/CutClean

The datasets commonly used to benchmark model debiasing are built with a strong correlation between the target and the private attribute. CutClean studies privacy leakage that arises independently of such spurious correlations, so every dataset here is a custom version balanced with respect to both the target and the private attribute.

Contents

File	Size	Content
corrupted_cifar10_unbiased.tar.gz	25 MB	12,700 images, ten target classes, ten corruption types as private attribute
waterbirds_unbiased.tar.gz	283 MB	3,324 images, waterbird/landbird target, background as private attribute
celeba_manifests/*.csv	1 MB	split manifests for the two CelebA configurations
SHA256SUMS		checksums of the archives

Both archives expand to one directory per split, containing one folder per target class. The label and the private attribute are encoded in each file name as ..._lbl<TARGET>_bias<PRIVATE>.png.

corrupted_cifar_unbiased/{train,valid,test}/<class>/img_XXXXXX_lbl<Y>_bias<B>.png
waterbirds_unbiased/{train,val,test}/<class>/img_XXXXXX_id<N>_lbl<Y>_bias<B>.png

Split sizes follow the paper: 8,900 / 2,500 / 1,300 for Corrupted-CIFAR10 and 2,328 / 664 / 332 for Waterbirds.

CelebA is not redistributed

The CelebA images are covered by a licence that restricts redistribution, so only the manifests are published here. Each CSV lists, for seeds 0, 1 and 2, the exact balanced subset used in the experiments:

column	meaning
seed	seed of the balanced subsampling
split	train, valid or test
image_id	file name inside img_align_celeba/
target	target attribute (Blond_Hair or Heavy_Makeup)
bias	private attribute, gender

Subset sizes match the paper: 5,548 / 728 / 720 for the blond-hair configuration and 812 / 36 / 88 for heavy make-up. The training code reproduces these subsets on its own from the official CelebA release, which it downloads on first use; the manifests are provided so the exact selection can be audited or pinned independently of the pandas version.

Usage

from huggingface_hub import snapshot_download
snapshot_download(repo_id="imDalton/cutclean-datasets", repo_type="dataset",
                  local_dir="data")

The reference implementation does this for you:

python src/data_setup.py --root ./data

Provenance and licensing

These are derived datasets. The terms of the original sources apply, and each of them must be consulted before use:

• Corrupted-CIFAR10 is built from CIFAR-10 (Krizhevsky, 2009) by applying the corruption functions of Hendrycks and Dietterich, Benchmarking Neural Network Robustness to Common Corruptions and Perturbations, ICLR 2019 (Apache-2.0), following the protocol of Nam et al., Learning from Failure, NeurIPS 2020.
• Waterbirds is built from CUB-200-2011 (Wah et al., 2011) and Places (Zhou et al., 2017), following Sagawa et al., Distributionally Robust Neural Networks, ICLR 2020. CUB-200-2011 is made available for non-commercial research purposes only, and that restriction carries over to this derivative.
• CelebA (Liu et al., ICCV 2015) is available for non-commercial research purposes only. No CelebA image is redistributed here.

Redistribution is limited to non-commercial research use. If you are a rights holder and consider that any material here should not be redistributed, please open a discussion on this repository and it will be removed.

Citation

@inproceedings{magliolo2026cutclean,
  author    = {Magliolo, Leonardo and Pastore, Vito Paolo and Valenzise, Giuseppe and Tartaglione, Enzo},
  title     = {CutClean: Neural Network Pruning for Privacy-Preserving Inference},
  booktitle = {Pattern Recognition -- 28th International Conference on Pattern Recognition, {ICPR} 2026, Lyon, France, August 17--22, 2026, Proceedings},
  series    = {Lecture Notes in Computer Science},
  publisher = {Springer Nature Switzerland},
  year      = {2026},
  pages     = {450--465},
  doi       = {10.1007/978-3-032-31452-9_30}
}