An evaluation benchmark released under CC BY-NC 4.0 for research and evaluation. Email danielbabalola@alumni.upenn.edu for commercial requests. Optional: tell me what you work on, and opt in below if you want a heads-up when datasets like this drop. I plan the next dataset around what people actually need.

KYC Passport Deepfake / Presentation-Attack Benchmark

HiDream-O1 · Qwen-Image-Edit · 23 document-reproduction conditions

A presentation-attack detection (PAD) evaluation set for identity-document face imagery: real reference faces and generator-attributed synthetic faces, each carried through 23 acquisition conditions that emulate how a passport photo actually reaches a KYC system: print and scan, photocopy, fax, phone recapture, JPEG chains, and background substitution.

159,804 images · 6,948 base media · 628 identities · 23 conditions · 12 demographic cells · 2 generators

This is the complete benchmark: every reference, every synthetic face, every condition, capped at 5 images per identity. Access is gated and approved manually.

2,316 real reference faces and 4,632 synthetic faces (two per reference, one from each generator), each rendered under 23 acquisition conditions.

Why this exists

Most deepfake-detection benchmarks score models on pristine images. Identity documents are almost never pristine. They are printed, scanned, photocopied, faxed, and re-photographed before a detector ever sees them, and that channel is where detectors fail.

On the full benchmark, macro-averaged AUC across 13 public detectors falls from 0.742 on clean images to ~0.528 under print/scan and photocopy, close to chance. Detector ranking also reorders under those conditions, so a model chosen on clean accuracy is not the model you would choose for the deployed channel.

That result is reproducible here: every condition and every image is present.

Field	Values
`label`	`0` = bona fide (real), `1` = synthetic
`kind`	`real`, `fake`
`condition`	`clean` + 19 emulations + 3 background strata (23 total)
`layer`	`clean`, `passport_emul`, `passport_bg`
`generator_model`	`hidream_o1`, `qwen_image_edit` (empty for reals)
`skin_tone`	`very_light`, `light`, `intermediate`, `tan`, `brown`, `dark`
`gender`	`male`, `female`
`sha256`	checksum of the delivered file

Conditions. clean; JPEG q30/q50/q70/q95; print_scan, photocopy, phone_scan_color, sensor_noise, demosaic, double_jpeg, downscale_id, grayscale_luma, jpeg2000, lens_blur; chains chain_print_scan, chain_fax, chain_flatbed_gray, chain_phone_bw_copy, chain_phone_color; backgrounds bg_plain_light, bg_gradient, bg_busy.

Perturbations are applied to the normalized image (JPEG q95, 768px longest side), never to raw source, so every condition differs from clean only by the emulated channel.

Provenance and licensing

Real faces are licensed from commercial stock-media providers. Synthetic faces are generated in house from those references. This dataset is released CC BY-NC 4.0, non-commercial research and evaluation only. Commercial use requires a separate license.

No images depict real identity documents. Synthetic faces are placed in passport-style framing; they are not reproductions of any genuine travel document, and no real document templates, security features, or holder data appear anywhere in the set.

Known limitations (read before publishing on this data)

1. Effective sample size is identities, not images. Every identity is capped at 5 images (enforced, verified: max = 5), but 2,316 reference images still come from only 628 identities. Identity-clustered confidence intervals are required; per-image CIs will be optimistic by roughly the square root of the clustering factor. Identity groupings ship on request. A one-image-per-identity subset (628 fully independent references) is available for work needing strict independence, though it cannot support within-identity analysis.

2. Cells are unevenly populated and reflect the source pool, not a target distribution: light_female is the largest at 37% of references, very_light_male the smallest at 0.4% (10 images across 5 identities). Small-cell results are indicative, not statistical. Report per-cell n alongside any per-cell metric.

3. Two generators only. hidream_o1 and qwen_image_edit. This is not a survey of generator families, and results should not be generalized to "diffusion editors."

4. One prompt, one seed per reference. Prompt homogeneity is not separated from generator signature.

What is not in this release

The images are complete. What ships separately, on request:

Per-condition and per-cell detector failure rates across the 13 public detectors we evaluated on exactly this corpus.
Identity groupings, needed for correctly clustered confidence intervals.
The one-image-per-identity subset for strict-independence work.
A sequestered set for independent evaluation of a detector you are shipping.
Custom generation matched to your generators, demographics, and conditions.

Tick the box in the access form or get in touch.

Citation

@misc{kyc_passport_pad_v1,
  title  = {A KYC Passport Deepfake and Presentation-Attack Detection Benchmark:
            HiDream-O1 and Qwen-Image-Edit Faces Under 23 Document-Reproduction
            Conditions},
  author = {Babalola, Daniel},
  year   = {2026},
  note   = {159,804 images, 6,948 media, 628 identities, 23 acquisition conditions.}
}

Intended use

Evaluating robustness of face/deepfake detectors under document-reproduction channels; measuring degradation between clean and acquisition-degraded conditions; fairness analysis as an exploratory signal.

Out of scope: producing deceptive identity documents, attacking live identity-verification systems, or any use implying these are genuine travel documents.

Downloads last month: 21

Total file size:

18.7 GB

Request access to the KYC Passport PAD Benchmark