FaizPalwala/SFHQ-VirtualID-Bench

SFHQ-VirtualID-Bench Dataset Card

Summary

SFHQ-VirtualID-Bench is a synthetic, identity-conditioned face dataset of 750 synthetic identities, built as a benchmark for machine unlearning. Each identity is a deletion unit: all 90 aligned 224×224 crops of one identity share a single identity-level split (retain or forget), and 75 identities follow a sequential 15-step forgetting protocol. Every identity contributes both per-image image_subset values (train + holdout), so retention and forgetting-generalisation can be measured on genuinely held-out images (MUFAC-aligned evaluation).

The release ships two artifacts:

This dataset is synthetic. Synthetic does not mean risk-free — residual likeness, demographic bias, and training-data memorisation are possible and must be acknowledged.

Composition

The imbalanced artifact ships 36,064 rows (11 fewer than 75 × 100): 123 candidates failed the preprocess quality gate (blur/unreadable), capping 10 high-bin identity pools at 98–99 crops.

Labels

• —Schedule columns are balanced-only. The balanced artifact ships both forget_step (uniform baseline — step index == cumulative forgotten count) and forget_step_poisson (seeded arrival-model stress test, λ=5 rebalanced to 75; evaluate by cumulative count, not step index). The imbalanced artifact deliberately carries no schedule columns — its only experimental axis is the popularity gradient.
• —Prompt metadata is stripped from Bench (pose/expression/lighting/ setting/camera ship only in the Raw release).
• —Age labels are per-image, not per-identity. Each crop inherits the gender/age estimate taken on its parent 1024×1024 candidate, so one identity can span several age groups. This is intentional — do not collapse to a per-identity constant.
• —Age-group labels are proxy estimates from the InsightFace classifier — not verified demographic attributes. The confound columns are provided so downstream MIA/fairness analyses can control for per-identity variation.

Split isolation invariant

Every identity_id maps to exactly one split (retain or forget), and every identity contributes both image_subset values (train + holdout). No identity's images appear in multiple splits. This invariant is enforced by the build step and validated by validate_release.py before publication.

Intended use

The 224×224 Bench is the primary machine-unlearning benchmark: aligned crops match the ImageNet training regime of downstream ResNet-18 classifiers, so pretrained features activate at full fidelity from epoch 1. The imbalanced variant adds a 5:1 popularity gradient for long-tail unlearning stress-testing, with a uniform 18-image holdout per identity so probe stability is comparable across all popularity tiers. Typical tasks: unlearning, membership-inference attacks, forgetting-generalisation, retention.

Method

Synthetic identities were generated with InstantID + Juggernaut-XL-v9 + ControlNet from CC0 synthetic SFHQ seed images (CLIP+KMeans-diverse seed selection); faces were detected, aligned, and cropped to 224×224, then quality-gated (sharpness + decodability) before split assignment. The arcface_similarity column is a confound control — the similarity floor was not applied as a filter in this release. Exact model revisions, generation configuration, and licence URLs are recorded in RELEASE_MANIFEST.json and THIRD_PARTY_NOTICES.md. No model weights are distributed.

Limitations

• —Synthetic, not anonymous. Generated faces may retain unintended resemblance to real persons through the training data of the generators.
• —Proxy labels. Age-group fields are model-estimated and carry classifier bias; they are not ground-truth demographics.
• —Demographic bias. The SFHQ source may over-/under-represent certain appearances; no fairness correction has been applied.
• —Artifacts. Diffusion artifacts (blur, distorted features) are most visible in 224×224 downscaled crops.
• —Imbalance is synthetic. The popularity gradient is a deliberate down-sampling, not real-world distribution.

Privacy and release policy

The release does not include: raw SFHQ source images, seed-to-output linkage, rejected candidates, ArcFace embeddings or biometric templates, or model weights. Users must comply with the upstream licence terms of the generation models (see THIRD_PARTY_NOTICES.md). A takedown process for credible likeness complaints is available via the repository.

Licence and citation

• —Code: MIT
• —Images and metadata: Non-commercial research use only, with a prohibited-use clause (no biometric identification, surveillance, authentication, impersonation, or high-impact decisions about people). See `LICENSE`.

@dataset{sfhq_virtualid_bench,
  title     = {{SFHQ-VirtualID-Bench}: Synthetic Identity-Conditioned Aligned
               Face Crops for Machine Unlearning},
  author    = {Faiz Palwala},
  year      = {2026},
  version   = {1.0.0},
  doi       = {10.5281/zenodo.21877893},
  url       = {https://github.com/FaizPalwala/virtual-id-gen},
}

See `CITATION.cff` for the complete metadata file.

Acknowledgements

This work was undertaken on the Aire HPC system at the University of Leeds, UK.