Spoof in the Wild Dataset
There are >10K videos from 3 types of 3D mask attacks tailored
for iBeta level 2 certification
Check samples on Kaggle
Introduction
The Face Spoofing in the Wild Multi-Mask dataset provides a practical benchmark for strengthening liveness detection systems under real-world conditions. It brings together three realistic mask-attack modalities – cardboard cutouts, fabric/printed textile masks, and 3D latex masks – captured across varied subjects, cameras, viewpoints, and lighting. This diversity makes it especially useful for assessing, stress-testing, and fine-tuning passive PAD models, an important step toward iBeta Level 2 readiness
The key value is high variety
Different actors, a large number of individual masks, various lighting (indoor/outdoor, natural/artificial light), backgrounds and locations, shooting angles and distances, different cameras/lenses, accessories (glasses/headwear/beards). This range reduces retraining on specific mask textures and increases generalizability in passive liveness detection and attack type recognition
Types of attacks
- Cardboard — thick cardboard with cut-out eyes Flat face print; eyes cut out to allow for real eye movement. Flat image with hard edges around the outline
- Textile — thin elastic mesh mask with face print Nylon/elastic fabric mask; full-color face print, fabric fits snugly around the head and partially replicates the 3D shape
- Latex — volume 3D latex masks Full-size realistic masks with pronounced “skin” texture. Sometimes they have cut-out eyes and are complemented by external attributes
Why in-the-wild collection improves robustness
Crowd-sourced, in-the-wild data covers the natural long-tail diversity of real scenarios, so models trained on it generalize better and rely less on hidden shortcuts typical of controlled, in-office datasets
How in-the-wild fundamentally differs from in-house
- Heterogeneous hardware. Smartphones/webcams across brands and generations → different sensors, optics, ISP pipelines, noise/exposure, stabilization, frame rates, and codecs
- Unstaged conditions. Random lighting (daylight/artificial, flicker, backlight), backgrounds, reflections, shadows, and outdoor/indoor scenes
- Human behavior. Natural poses, micro-movements, expressions, speed and amplitude of head turns, varying camera distance
- Broad spectrum of PAIs/masks. Different materials, shapes, and application methods; real-world artifacts (glare, folds, misalignment); plus “imperfect” spoofing attempts
- Demographic and cultural diversity. Skin tones, makeup, accessories (glasses, headwear), styles—rarely covered in office setups
In in-house collected datasets, even with artificial variation of backgrounds/angles, common constants remain: the same camera pool, typical lighting and backgrounds, repeated rooms and collection crews. Models quickly latch onto these spurious cues (e.g., characteristic white balance or wall texture), which harms transfer to real-world conditions
Because the sources of variability differ, in-the-wild and in-house datasets have weak overlap. This makes our dataset a valuable external test bed: if a model performs well here, the likelihood of failure in real production is substantially lower
File format and accessibility
- Format: Videos are optimized for compatibility with mainstream ML frameworks
- Resolution and frame rate: Videos are high-resolution with frame rates calibrated for capturing quick and realistic mask placements, ensuring precise data for model training
Legal & Compliance
We prioritize data privacy, ethical AI development, and regulatory compliance. Our Silicone Mask Attack Dataset is collected and processed in full accordance with global data protection standards including GDPR, ensuring legality, security, and responsible AI practices
Sample dataset
A sample version of this dataset is available on Kaggle. Leave a request for additional samples in the form below
Have a question?
We collect data from our internal team. All information is further verified by our specialists
Once your enquiry has been sent, we will contact you to discuss the details and complete the necessary paperwork. The timing of receiving the dataset depends on the specific request and additional requirements
Our unique selling point is to provide legally clean datasets to our customers. We obtain the consent from all the participants to use their data for AI model development. We are able to provide comprensive reporting on the licensing, data collection and privacy compliance of our datasets. Although there seems to be a diverse response to how to control AI development and deployment, we are able to service global customers seeking to launch global AI products.
The dataset follows iBeta testing protocols and includes diverse attack scenarios that mirror real-world spoofing attempts. It covers both passive and active liveness testing requirements with proper demographic representation and standardized capture conditions essential for certification preparation
The price depends on your specific requirements. Please submit a request to receive a free consultation
Contact us
Tell us about yourself, and get access to free samples of the dataset
Didn't find what you were looking for?
Our collection includes many datasets for various requests
iBeta Level 1 Dataset
– 22,000+ videos
– 80+ participants
– zoom in and
zoom out
iBeta Level 2 Dataset
– 25 000+ videos
– 3D masks
– iBeta Level 2
Replay Display attacks
– 5,000+ videos
– 1,000+ participants
– Balanced mix of genders and ethnicities
Photo Print Dataset
– 7000+ videos.
– 10-20 second each video
– Mix of genders
