How Local Liveness Checks Work in Facenox

Face recognition asks whether a face matches an enrolled member. Liveness checks address a different question: whether the camera input appears to come from a live person rather than a presentation such as a photograph or screen.

Facenox runs these checks on the attendance computer. Cloud is not required for local inference.

Correction, October 23, 2026: This article now describes AuraFAS, replacing the earlier MiniFASNetV2 description. The previously published attack rates, timing figures, and protection guarantees have been withdrawn because reproducible supporting results were not available.

What the local model receives

The current liveness pipeline uses AuraFAS through ONNX Runtime. Preprocessing supplies a tight face crop, a wider context crop, and a frequency-domain input derived using an FFT. The implementation combines model outputs with detection validation and tracked liveness state.

What an administrator should test

Use the computer, camera, placement, and lighting intended for daily operation. Check ordinary enrollment and check-in behavior with representative users before rollout.

Evaluate both mistakes: an unwanted input being accepted and a legitimate person being rejected. If conducting spoof tests, record the test method, model version, settings, number of attempts, and outcomes. Obtain appropriate permission and protect the test data.

Changing thresholds can change the balance between acceptance and rejection. A faster or more permissive setup is not automatically a better one.

Accuracy and attendance review

Liveness checks can make mistakes and do not guarantee detection of every spoof attempt. Review unusual or disputed attendance records, and protect access to the computer running the checks.

When comparing test results, check the hardware, software versions, attack types, thresholds, and number of attempts. A result from one setup may not reflect another.

Implementation references

See the liveness implementation and AuraFAS input and inference tests.