Security outcomes
Attack acceptance and genuine-user rejection, separated by attack type and evaluation protocol. PAD-layer APCER and BPCER are reported with their applicable protocol.
05EVALUATION PLAN
Evidence before performance claims.
The design defines a controlled comparison of fusion strategies. Measured accuracy, latency and throughput results have not yet been published.
Current status: planned evaluation. No percentages or illustrative scores on this site should be read as experimental results.
01COMPARATIVE EXPERIMENTS
The same evidence. Different policies.
Subject-disjoint evaluation and fixed validation/test splits are planned. C, C-Q and D are compared using the same module outputs and mandatory safety gates so the effect of quality and attack rules can be assessed separately.
| CONFIGURATION | PURPOSE | ROLE |
|---|---|---|
| A · Face only | Identity-only diagnostic control | Offline baseline |
| B · Face + PAD | Measure the contribution of visual PAD | Offline baseline |
| C · L1 | Fixed-weight full-channel fusion | Shared gates |
| C-Q | Quality-weighted full-channel fusion | Shared gates |
| D · L2 | Quality- and attack-aware fusion | Shared gates + pattern rules |
| L3 | ML-assisted advisory model | Offline evaluation |
02WHAT WILL BE REPORTED
Security, usability and cost—together.
Coverage & retries
ACCEPT, REJECT and RETRY distributions for genuine and attack attempts, definitive-decision coverage and technical failures.
Runtime behaviour
Latency and resource usage under declared capture and deployment profiles. Short-window rPPG uncertainty is reported explicitly.
Physical presentation attacks and digital input injection are evaluated separately. ISO/IEC 30107-3 terminology is a design reference; the project does not claim certification or full standards conformance.
FROM DESIGN TO IMPLEMENTATION
Explore the thinking behind BioAuth.
Read the architecture, evidence contracts and experimental design.