Detect Face Liveness (MiniFASNet)

@deepinsight1 VersionDetect

Scores each Rectangle<Double> face box in an Image for passive presentation-attack detection with a dual MiniFASNet anti-spoofing ensemble, emitting one liveness Classification per face -- ``live`` / ``print`` / ``replay`` plus confidence -- in box order.

How it fits

Typical backends

Detect faces in a selfie, score each face for liveness, and ship the per-face verdicts as JSON for an identity-validation check.

Keep only genuine-live faces from a live camera by gating the liveness verdict through a filter before downstream face matching.

Caveats

I/O contractOne Classification is emitted per input face box, in input box order, so an empty box list yields an empty output without running the network and output cardinality always matches the input box list.
LimitationThe label is a native three-class verdict -- ``live``, ``print`` (2D print/paper attack), or ``replay`` (screen/video replay attack); the robust signal is live versus spoof, and the print/replay split follows the upstream MiniFASNet class-index convention rather than an independently calibrated attack-type classifier.
Hard constraintThe weights directory referenced by model is scanned recursively for the MiniFASNet ONNX checkpoints, whose architecture names recover the per-model crop scale; a checkpoint whose filename carries no recognised architecture name aborts startup.
CostBoth ensemble models run once per face box, so per-frame cost grows linearly with the number of faces in the frame.
FallbackRequesting a GPU via device when none is visible falls back to CPU with a warning and continues at CPU speeds.
State lifetimeThe model weights and device are captured once at startup, so runtime changes take effect only after a redeploy.

Versions

ca1d6331latestdefault
Automated release
6/28/2026