UtilFace is released for the evaluation of face de-identification methods. By requesting access you agree to (1) use it for non-commercial research only, (2) not use it to train face recognition systems or to re-identify any individual, (3) obtain the underlying WebFace42M images only from their official providers under the WebFace42M terms, and (4) honour removal requests propagated through updates to this repository.

UtilFace

UtilFace is the benchmark set of the NeurIPS 2026 Evaluations & Datasets track paper How Private is Private? A Comparative Study for Face De-Identification (Hui Wei, Hao Yu, Hui Kuurila-Zhang, Guoying Zhao; Center for Machine Vision and Signal Analysis, University of Oulu). It is a curated, demographically balanced set of 99,928 face images of 2,069 identities used, together with the HiFD metric, to evaluate face de-identification (FDeID) methods.

This repository contains no images. It releases the identity keys and file list that define UtilFace inside WebFace42M, the sampling strata, the HiFD scores of the 12 baseline methods, and mini-UtilFace, a 2,000-image validation bundle of per-image sub-scores that reproduces the published ranking without any face data.

Code (HiFD toolkit and estimator pipeline): https://github.com/infraface/hifd
Croissant metadata (core + Responsible AI): croissant.json

path	rows	description
`data/images.csv`	99,928	one row per benchmark image
`data/identities.csv`	2,069	one row per identity
`scores/hifd_scores.csv`	12	HiFD components and composites per de-identification method (paper Tab. 3)
`mini_utilface/`	2,000	validation bundle: `subscores.npz`, `keys.csv`, `video_axes.json`, `published_scores.csv`, `verify.py`, `README.md`
`croissant.json`		Croissant 1.1 metadata with RAI and provenance fields

`data/images.csv`

column	type	meaning
`image_index`	int	stable index of the image in the UtilFace build (used as `lmdb_key` in `mini_utilface/`)
`image`	str	path of the image inside WebFace42M, `<identity>/<file>.jpg`
`identity`	str	WebFace42M identity folder
`scored`	0/1	1 if the image enters the HiFD tables (99,365 images; the remaining 563 have no output from the face-embedding stage on the original image)
`in_mini_utilface`	0/1	1 if the image is one of the 2,000 mini-UtilFace images

`data/identities.csv`

column	type	meaning
`identity`	str	WebFace42M identity folder
`num_images`	int	images of this identity in UtilFace (5 to 510)
`gender_stratum`	str	sampling stratum: majority vote of the gender estimator over the identity's images
`ethnicity_stratum`	str	sampling stratum: majority vote of the FairFace 7-class estimator over the identity's images

Reconstructing the images

Obtain WebFace42M from its official providers (https://www.face-benchmark.org) under the WebFace42M terms of use.
Select the files listed in data/images.csv (image column).
Apply the UtilFace preprocessing (blind face restoration to 512x512 and the LMDB packing).

Curation

Four public face-recognition corpora were collected: WebFace42M, Glint360K, MS1MV2 and MegaFace. After cleaning (removal of images with zero or multiple faces, non-RGB and heavily blurred images, and identity-consistency checks), the samples that met the quality and identity-consistency requirements all come from WebFace42M, so every UtilFace image is a WebFace42M image.

Pipeline, in order:

Clean WebFace42M as above. Restore low-resolution faces (GFPGAN or CodeFormer, selected by pose; upsampling capped at 4x) to 448x448 or 512x512.
Keep images with NIQE < 5.0 and resize them to 512x512 (8.69M images).
Annotate every image with RetinaFace (detection), MiVOLO (age, gender), FairFace (ethnicity), POSTER (expression), PIPNet (landmarks) and L2CS-Net (gaze).
Drop no-face and multi-face images and identities with fewer than five images. Assign each identity a gender-by-ethnicity stratum by majority vote, rank identities within each stratum by image count and attribute diversity (expression and pose entropy), and select identities until each of the 14 cells holds about 7,100 images. All images of a selected identity are kept.

Responsible AI

Personal and sensitive information. The underlying images are photographs of real people (largely public figures) collected from the web by the WebFace260M authors without individual consent. This release contains no images, face embeddings, or per-image predicted attributes. The identity keys are opaque WebFace42M folder names and identify a person only to someone who already holds WebFace42M. The identity-level strata are machine-inferred gender and ethnicity labels, which are sensitive. They are released only as sampling labels.

Biases. Balance is enforced on predicted gender (binary) and ethnicity (7 FairFace classes), so it inherits the errors and category choices of those estimators. Age, expression and pose are balanced only through the diversity ranking. The source is web imagery of public figures, which skews toward frontal, well-lit, adult faces. Blind face restoration can homogenise skin texture and other fine detail. The human study rates unmodified UtilFace images 3.32/5 on naturalness.

Limitations. UtilFace is an image-level benchmark. HiFD's L2 micro-expression and L3 rPPG components are measured on DFME and PURE/MMPD, which are not part of this release. Rebuilding pixel-identical images requires WebFace42M access and the preprocessing scripts. Strata are estimator outputs and are wrong for some identities.

Social impact. UtilFace provides a common, annotation-free yardstick for privacy-preserving face analysis and exposes silent failures (for example, destroyed physiological signals) before deployment. Foreseeable misuse includes tuning methods to evade lawful face recognition. Mitigations are the release design (no pixels, no embeddings, gated research-only terms) and honouring removal requests.

Maintenance and removal. Maintained by the authors. To request the removal of an identity, open a discussion on this repository or e-mail the corresponding author. Removed identities are deleted from all files in a new version, and earlier versions are noted in the changelog below.

License

The files in this repository are released under CC BY-NC 4.0. The underlying WebFace42M images are not covered by this license and remain subject to the WebFace42M terms of use. The HiFD code is MIT-licensed (see the code repository).

Citation

@inproceedings{wei2026private,
  title     = {How Private is Private? A Comparative Study for Face De-Identification},
  author    = {Wei, Hui and Yu, Hao and Kuurila-Zhang, Hui and Zhao, Guoying},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Please also cite WebFace260M/WebFace42M (Zhu et al., CVPR 2021).

Changelog

v1.0 (2026-10): initial public release for the NeurIPS 2026 camera-ready.

Downloads last month: -

Total file size:

6.07 MB

Research-only access to UtilFace