AxonData/2d-paper-mask-face-anti-spoofing

---
license: cc-by-nc-4.0
task_categories:
- video-classification
- image-classification
tags:
- face-anti-spoofing
- liveness-detection
- paper-mask
- print-attack
- presentation-attack-detection
- face-spoofing
- cut-out-mask
- 2d-mask-attack
- biometrics
- PAD
- computer-vision
size_categories:
- 1K<n<10K
language:
- en
---
# Cut-Out Paper Mask Face Spoofing Dataset

3,000 videos of partial 2D paper mask attacks from 50 participants, recorded on Galaxy A54 and iPhone 14 Pro. Built for training and evaluating **face anti-spoofing**, **liveness detection**, and **presentation attack detection** systems.


> **Full dataset for commercial use** — request a license at [axonlab.ai](https://axonlab.ai/?utm_source=hugging-face&utm_medium=referral&utm_campaign=paper-mask-dataset)

## Why Cut-Out Masks Are a Harder Problem

Standard **print attack** and **paper mask** datasets use full-face photo printouts. Cut-out masks are fundamentally different:

- The attacker's **real skin, eyes, or mouth remain visible** — bypassing texture-based detectors
- **Real facial movement** (blinking, lip movement) co-exists with the printed region
- **Edge artifacts** between real skin and paper are subtle and localized

This makes cut-out paper mask attacks a realistic blind spot for PAD systems trained only on full-face 2D or 3D attacks

## Dataset Specifications

- **3,000 videos** from **50 unique participants**
- **Multiple mask variations** — printed cut-outs covering different facial regions (eyes, mouth, nose area) and their combinations
- **Dual-device**: Galaxy A54 (Android) + iPhone 14 Pro (iOS)
- **15 seconds per video** with active liveness features: zoom-in/out, natural head movements, blinking
- **Varied environments**: diverse real-world backgrounds and lighting conditions
- Masks printed on high-quality paper, skin-tone matched, attached with transparent tape or held by hand

## Applications

- **Face anti-spoofing** — train PAD models specifically for partial paper overlay attacks
- **Liveness detection** — improve robustness beyond replay and full-face print attacks
- **iBeta certification preparation** — test against realistic 2D attack scenarios before Level 1/2 submission
- **Print attack detection** — extend existing models to handle cut-out overlays

## Need More Data?

This dataset is a ready-made sample. We offer **custom data collection** for cut-out paper mask attacks tailored to your requirements — including larger participant pools, additional devices, specific demographic distributions, and custom mask configurations

Contact us at [axonlab.ai](https://axonlab.ai/?utm_source=hugging-face&utm_medium=referral&utm_campaign=paper-mask-dataset-custom) to discuss your project