Build and improve detection
Evaluate and retrain your own models on what your generator does not produce.
- Region labels
- Held-out test split
- Dated additions
By IndustriesBy Industries- By Use Cases
By Use Cases
New: labelled deepfake datasets for ID and selfie detection.
Get a free sample
Calculate the impact of deepfake prevention instantly using our new Calculator!
Calculate Now
How bunq Secured KYC Onboarding and Cut 600% in Manual Review Time
See Case Study
Learn How Certta Stopped Synthetic Identities at Scale With Deepfake Detection
Learn More
Book a Demo & See Deepfake Detection in Your Flow
See our deepfake detection flag a fake selfie, document, or voice clone inside a live identity verification flow. Book 30 minutes and watch it run on your own KYC data, in seconds.
Deepfake datasets for training and evaluation
Deepfake liveness datasets for training and evaluation
Would your system catch today's deepfakes?
Test your defences against evolving AI-generated and manipulated documents. Get the data to identify gaps, retrain your models and revalidate as threats change.
Test your liveness defences against evolving face swaps, replays and injection attacks. Get the data to identify gaps, retrain your models and revalidate as threats change.
11countries covered
Sep 2026newest design
Datedevery addition
5attack families
Aug 2026newest generator
Datedevery addition
NL_ID_0231_F · front · labelledmanifest row 4 of 5
Koninkrijk der NederlandenIdentiteitskaart · NLD
SurnameVAN DER BERG
Given namesANNA MARIA
NationalityNederlandse
Date of birth12 JUL 1991
Document no.NX3L2K7T9
Expiry15 FEB 2032
I<NLDNX3L2K7T94<<<<<<<<<<<<<<<
9107126F3202154NLD<<<<<<<<<<<4
VAN<DER<BERG<<ANNA<MARIA<<<<<<<
9107126F3202154NLD<<<<<<<<<<<4
VAN<DER<BERG<<ANNA<MARIA<<<<<<<
portrait_replacement
document_number_edit
text_field_edit
template_regeneration
SELF_0412_S · front camera · labelledmanifest row 2 of 25
Liveness checkpassive
face_swap
reenactment
screen_replay
injection_attack · virtual camera
fully_synthetic · diffusion
iPhone 15 · front · 1080×1440indoor · backlit
Additional coverage
Add coverage beyond your own datasets.
Explore attack methods and document variations your existing data may not cover. Dates show when each item was added to the datasets. Pick one to get that sample.
Planned · Q4 2026
Q4 2025Q1 2026Q2 2026Q3 2026next
Planned · Q4 2026
Q4 2025Q1 2026Q2 2026Q3 2026next
7 documents · 2 attack methods added since Q4 20254 attack methods added since Q4 2025New document coverageNew attack methodPlanned, not yet delivered
Why it matters
Your defences are only as current as the data behind them. Test them against today's deepfakes.
Test your current protection
Assess the verification system or provider you already use, before fraud does.
- Documents and liveness
- Labelled to the pixel
- No call needed
Add coverage, not duplication. You keep your own data and generators. We supply what they do not cover yet, under a perpetual licence for internal training and QA.
Sample data
See what you're testing.
Hover a field to see what it points at. Boxes are normalised per side, so a reviewer can check the label against the pixels in a minute.
Sample data
See what you're testing.
Attack family, generator, spoof medium, capture device and challenge type, on every still and every frame range. Hover a field to see what it points at.
Koninkrijk der NederlandenIdentiteitskaart · NLD
SurnameVAN DER BERG
Given namesANNA MARIA
NationalityNederlandse
Date of birth12 JUL 1991
Document no.NX3L2K7T9
Expiry15 FEB 2032
I<NLDNX3L2K7T94<<<<<<<<<<<<<<<
9107126F3202154NLD<<<<<<<<<<<4
VAN<DER<BERG<<ANNA<MARIA<<<<<<<
9107126F3202154NLD<<<<<<<<<<<4
VAN<DER<BERG<<ANNA<MARIA<<<<<<<
region_bbox 0.05 0.23 0.28 0.56
doc_type · country · template_id
image_id · side · hash
family · technique
label · split
image_idNL_ID_0231_F
hashsha256:9f3a…
sidefront
countryNLD
doc_typeid_card
labelfake
familymanipulated
techniqueportrait_replacement
region_bbox0.05 0.23 0.28 0.56
template_idNL_ID_2021
splittest · sealed
● teal: pixel-level fields○ white: template and provenance fieldsOverlays ship as separate debug images
Liveness checkblink
region_bbox 0.31 0.18 0.38 0.46
spoof_medium · capture_device · lighting
image_id · modality · frame_range
challenge · label · split
family · generator_id
Pixel 8 · front · 1080×1440outdoor · low light
image_idSELF_0412_S
modalityvideo · 3 s · 30 fps
capture_deviceandroid · pixel_8 · front
challengeblink · active
lightingoutdoor · low_light
labelattack
familyface_swap
generator_idGEN_2026_08_diff_v3
region_bbox0.31 0.18 0.38 0.46
spoof_mediumnone · digital
splittest · sealed
frame_range12–84 · attack visible
● teal: pixel and frame-level fields○ white: capture, medium and generator fieldsVideo packs carry one row per attacked frame range
deliverySFTP · versioned
pack_2026Q3_v1.2/
NLD/
front/ 124 jpg
back/ 124 jpg
USA/
front/ 96 jpg
back/ 96 jpg
ARG/ BRA/ GBR/ … +6
overlays/ region debug png
manifest.csv 1 row per image
README.md attack categories · hygiene
CHANGELOG.md
test_sealed_v1.2/ delivered separately
EU hosting · SHA-256 per file · no watermarks, no recompression
imagesphone-range JPEG
1,500 to 4,000 px long side
EXIF · stripped
- Format
- JPEG, q 75 to 95
- Colour
- sRGB
- Framing
- Full document, edges visible
- Surface
- No glare, no recompression
unit1 case = 1 identity
synthetic
identity
identity
=
ID card · front + back
passport · 1 image
Priced per case, never per image.
splitstrain · sealed test
disjoint templatesdisjoint identitiesdelivered separately
Train on one, measure on the other. Your number means something.
familiestwo, labelled
~60%
~40%
labelsper image, per side
labelfamilytechniqueregion_bboxtemplate_idsplit
Region overlaysseparate debug PNGs, one per labelled side, so a reviewer checks the box against the pixels without parsing anything
not includedon purpose
×
Where the data comes from
A deepfake detection company first.
The engine that trains our detectors, opened to yours.
DuckDuckGoose ships deepfake detection to banks and identity providers. Threat intelligence from where techniques are traded, plus the commercial, open-source and in-house generators we track, decide what enters the datasets next.
Sources
Production trafficAttacks seen against our detectors at banks and IDV providers.
Trading communitiesWhere techniques are sold before detectors see them.
GeneratorsCommercial, open-source and our own models, selected per technique.
Outputs
Our detectorsPhocus, DeepDetector and Waver, in production.
Dataset packsFor your detectors, with the same labels and hygiene as ours.
How to start
Start with samples. Build around your requirements.
The dataset is scoped to your requirements and budget. A call is optional at every stage. NDA, procurement and delivery details are in the FAQ.
01
Explore samples
Work email, terms ticked. A one-time link to the free samples. Check realism, labels and fit with your pipeline.
Get sample data →
02
Agree your dataset
Tell us your markets, attack types and budget. We write the scope and quotation. A call only if you want one.
Discuss your requirements →
03
Receive and test
SFTP delivery, versioned, with a held-out test split. Run it against your system and tell us what you found.
YouBothUs·Start with step one. It costs two minutes and a work email.
FAQ
The practical questions, answered.
Which countries and document types are covered today?
Most of the world. Passports, ID cards, residence permits, driving licences and visas, in a proprietary dataset that is among the largest of its kind, with new designs added as they appear. See the latest additions, or get sample data for a market you care about.
Can you add a country, document or attack method we need?
Yes. Two countries of your choice are added inside any evaluation. Other requests are scoped in writing before you commit, and planned additions stay labelled as planned until they are delivered. Tell us your requirements and we write the scope within two working days.
How current is the data?
Every addition is dated on the coverage timeline. New document designs and attack methods are added as they appear and labelled to the region or the face. The changelog records each addition.
Is there a face swap and liveness video dataset as well?
Yes. The selfie and liveness packs cover presentation and injection attacks: face swaps, reenactment, screen replay and injection, as stills and 3 s clips labelled to the face, across devices, lighting, challenge types and skin tones. Faces are consented or synthetic. Request the sample clips and run them against your liveness SDK.
Do we need to share genuine documents or customer data?
No. The packs are synthetic and you run them inside your own environment. Nothing from your pipeline is sent to us, and no genuine documents are needed to evaluate. See how an evaluation runs.
Does the dataset contain real personal data, and is it GDPR compliant?
No real personal data is delivered. Every identity is synthetic; portraits and data fields are replaced and no real person's data remains. Each pack ships with a provenance statement covering source templates, replacement method and lawful basis, written for your DPO.
Can we buy the dataset under a commercial licence for training?
Yes. Packs are licensed perpetually for internal training, testing and QA. Public corpora such as FaceForensics++ are research-only, which is one reason teams come to us. Licence and delivery details are in the technical specifications.
How is a dataset priced?
Per case rather than per image, so both sides of an ID card count as one case. A focused evaluation set is scoped to your requirements and budget and quoted in writing within two working days. Discuss your requirements to get a quotation.
How quickly can a dataset be delivered?
About three business weeks after the scope is agreed. Delivery is by SFTP, versioned, with one folder per country and side and a held-out test split delivered separately. See what lands on your SFTP.
Start with samples
What would you like to put to the test?
Get sample data now, or tell us your requirements and we write the scope within two working days. A call only if you want one.
