Verisoul Docs
Search
⌃K

Security and Accuracy

Security

FacePrint

Risk
Security Functionality
2D Images, Print Outs
3D FaceScan detects points in three-dimensions, meaning 2D images and prints are quickly dismissed
3D Masks, Ultra-realistic wax sculptures
A) 3D Liveness detection checks for liveness context throughout user’s face over time, meaning static or partially static faces (partially altered with wax/mask) will be detected
B) Active movement ensures user liveness; additionally, multiple user distances from camera generates high-fidelity liveness data
Video injection & DeepFakes
Technology detects when camera feed is being altered or user is trying to inject video
Additionally, technology detects deepfake videos & images
FaceScan alteration
FaceScans are encrypted to prevent alteration on the client side
Client-side device risk
A) SDK checking for risk signals on the device that would indicate likely fraud
B) Use obfuscation and checksums to ensure code base is not tampered with

VoicePrint

Risk
Security Functionality
Computer Generated Voices / DeepFakes
AI model detects artifacts of deepfake voices and microphone / injected audio - even when such artifacts are indistinguishable to the human ear
Replay attacks
Verisoul’s onboarding ensures random phrase and detects recorded replays
Multi-language
Voice vectorization is language independent and text independent - so users that speak multiple languages can be identified in all of them as the same person.

Accuracy

Definitions - imagine User A is already enrolled
False Acceptance Rate (FAR) - this is the probability that a given user B can pretend to be User A
  • This is the value that matters for authentication (getting into an existing account)
False Rejection Rate (FRR) - this is the probability that User A will be rejected when trying to authenticate again
  • This is the value that matters for uniqueness (preventing an existing user from falsely creating a duplicate / new account)
Accuracy:
  • FAR: 3 / 10,000,000,000 chance (Apple’s touch ID is 1/50K, and FaceID is 1/1M)
  • FRR: < 3 / 100,000
  • Works with beards, transparent glasses (not sunglasses), and makeup
  • 3-Dimensional modeling based on facial features results in skin-tone agnostic accuracy
Combination of 3d face vector over time + voice print creates extremely high fidelity solution.