iBeta Level 3 Dataset – High-Fidelity Mask Attacks

High-fidelity rubber mask attacks are 3D presentation attacks using professionally manufactured rubber face masks designed to replicate specific target identities with photorealistic detail. Unlike paper-based or cloth mask attacks, rubber masks produce realistic skin texture, depth, and three-dimensional face geometry that challenges even advanced liveness detection models. These attacks are the primary class tested in iBeta Level 3 PAD certification - the highest tier of biometric presentation attack detection testing

The dataset contains 1,000+ videos across 7 high-fidelity rubber masks, each custom-manufactured to replicate a specific target identity. Every mask is recorded with multiple attribute combinations: different hairstyles, glasses, and accessories, to maximize within-mask visual diversity. Videos are captured with front-facing (selfie) cameras on iOS and Android devices, include active liveness sequences (head movements, blinking, zoom-in/out), and span indoor office backgrounds under varied lighting

The dataset targets iBeta Level 3 PAD certification - the highest tier of biometric presentation attack detection testing. iBeta Level 3 tests biometric systems against ISO/IEC 30107-3 at the highest assurance level, where APCER, BPCER, and ACER thresholds are tighter than at Levels 1 and 2. This tier is required for deployments where commercial-grade 3D masks must be reliably detected

This dataset focuses specifically on rubber mask attacks, one of two mask families tested at iBeta Level 3. For custom 3D-printed resin masks, see the 3D Resin Mask Dataset. For combined rubber and resin coverage in a single training set, see the main iBeta Level 3 Dataset. Rubber and resin masks produce fundamentally different visual artifacts, so training on both is recommended for production models targeting L3 certification

Rubber mask attacks were captured using front-facing (selfie) cameras on iOS and Android device models. Each of the 7 masks is recorded across multiple attribute combinations: different hairstyles, glasses, and accessories, to maximize within-mask visual diversity. Every video follows an active liveness protocol with head movements, blinking, and zoom-in/out phases, captured across varied indoor office backgrounds under natural and artificial lighting conditions

The dataset is built for the most demanding iBeta PAD tier through two quality dimensions. First, professional mask construction: each of the 7 rubber masks is manufactured to iBeta Level 3 certification realism standards, replicating specific target identities with photorealistic skin texture and three-dimensional face geometry. Second, deep per-mask coverage: each mask is recorded across multiple attribute combinations and capture conditions, producing ~140 videos per mask for robust within-mask training. Axon Labs datasets have contributed to 21% of companies passing iBeta certification in 2025

Yes. All data is collected with explicit written informed consent from every participant in compliance with GDPR Article 9, which specifically governs the processing of biometric data. The dataset is licensed for commercial use in AI model training, validation, and iBeta certification preparation. Comprehensive compliance documentation, including consent provenance, collection methodology, and legal basis review, is available upon request

Yes. A sample version of this dataset is available on request, you can verify mask construction quality, active liveness sequences, per-mask attribute variation, and format compatibility with your training pipeline before committing to the full dataset. Submit a request through the form on this page for sample access, typically delivered within 1–2 business days