Intro

To our knowledge we’re the only company in the world that can perform 3D liveness check and identity concealment detection from a single 2D image. We outperform the competition (FaceTEC, BioID, Onfido, Huawei…) in speed and accuracy. Our implementation is Passive/Frictionless and only takes few milliseconds.

Identity concealment detects when a user tries to partially hide his/her face (e.g. 3D realistic mask, dark glasses…) or alter the facial features (e.g. heavy makeup, fake nose, fake beard…) to impersonate another user.

A facial recognition system without liveness detector is just useless.

We can detect and block all known spoofing attacks: Paper Print, Screen, Video Replay, 3D (silicone, paper, tissue...) realistic face mask, 2D paper mask, Concealment, Deepfake...

Next video shows a stress test on our implementation using different type of 2D/3D attacks:

---

The next video shows the result on a Youtube video from https://www.youtube.com/watch?v=2tgqX5WVhr0 to detect deepfakes:

---

Our passive (frictionless) face liveness detector uses SOTA (State Of The Art) deep learning techniques and can be freely tested with your own images at https://www.doubango.org/webapps/face-liveness/

• Github repo: https://github.com/DoubangoTelecom/FaceLivenessDetection-SDK

• Cloud-based implementation: https://www.doubango.org/webapps/face-liveness/

• Open source Computer Vision library: https://github.com/DoubangoTelecom/compv

Contents:

• C++ API
• Supported attacks
• Identity concealment
  • Type of concealment
• Deepfake detection
• Accuracy
  • The competition
• Testing the competition
  • FaceTEC
    • <1% FRR
    • 1/12,800,000 FAR
    • Liveness detection in 2 seconds
    • $100,000 Spoof Bounty Program
    • 800% improvements
  • BioID
  • Huawei
  • Wiko
  • Samsung
  • Sybrin
  • Onfido
  • Jumio
  • Minivision-ai (turingtech.vip)
• ISO ISO/IEC 30107
• Why liveness detector is required for facial recognition systems?
• Passive versus Active liveness detection
  • Passive liveness
  • Active liveness
  • Why passive is better?
    • Accuracy
    • Faster enrollment
    • Frictionless
    • Harder to break
    • Easier to integrate to an existing application
    • Less bandwidth and CPU usage
• Recommendations
• Improving the recall score
  • The 7 commandments
    • 1. Face must be large and close
    • 2. Do not trust online images
    • 3. Selfie-type image (front facing) with neutral expression
    • 4. Use high resolution images
    • 5. Do not use grayscale images
    • 6. Use reasonable compression ratio
    • 7. Disable “Beauty” filter
• Integration with your existing application
• Error and success codes
• Memory management design
  • Memory pooling
  • Minimal cache eviction
  • Aligned on SIMD and cache line size
  • Cache blocking
• Muti-threading design
• Known issues
  • Masks
  • Very dark skin
  • 3D realistic masks