Balanced — mixed-30-v2¶
The recommended default. Analyzes 30 frames (1 second of video) for the best overall accuracy with minimal latency.
At a Glance¶
| Model ID | mixed-30-v2 |
| Frames required | 30 |
| Capture time | ~1 second at 30 FPS |
| Avg response time | ~3s |
| Security level | Strongest |
When to Use¶
- Standard KYC / identity verification — the best default for most use cases
- Account creation where you need solid protection without excessive friction
- Financial services where regulatory requirements demand robust liveness checks
- Any flow where 1 second of capture is acceptable
Performance¶
Based on evaluation across 1,960 test videos (9 attack types, held out from training):
| EER | 4.0% |
| AUC | 0.991 |
| Balanced accuracy | 95.7% |
| Avg response (fast-check) | ~3s |
Per-Attack Detection¶
| Attack Type | Detection Rate |
|---|---|
| Physical spoofing | 98.3% |
| Deepfakes | 95.6% |
| Screen replay | 92.6% |
| Faceswap | 80.3% |
Operating Points¶
The tradeoff between security and user friction at different thresholds:
| Max BPCER (user rejection) | APCER (attack pass-through) |
|---|---|
| 1% | 12.8% |
| 5% | 3.7% |
| 10% | 1.6% |
| Max APCER (attack pass-through) | BPCER (user rejection) |
|---|---|
| 1% | 10.6% |
| 5% | 3.1% |
| 10% | 1.2% |
What this means
At the EER operating point, 96% of attacks are correctly blocked and 96% of legitimate users pass. The 30-frame model delivers the best accuracy across all variants — the recommended choice for production.
Tradeoffs¶
- Best overall accuracy — 4.0% EER and 0.991 AUC, the top performer across all variants
- Strong across all attack types — 95%+ detection for deepfakes and physical spoofing
- Only 1 second of capture — acceptable for virtually all verification flows
- Best calibration — confidence scores most accurately reflect true probabilities (lowest ECE and Brier score)