litert-community/Silent-Face-Anti-Spoofing-LiteRT

Silent-Face Anti-Spoofing (MiniFASNetV2) — Face liveness (LiteRT GPU)

On-device face liveness / anti-spoofing running fully on the LiteRT CompiledModel GPU delegate (no CPU fallback). Silent-Face-Anti-Spoofing detects presentation attacks — a printed photo or a replayed screen shown to the camera — so a live face passes and a fake is rejected. The anti-fraud building block for face login / e-KYC. Tiny (1.85 MB), ~5 ms/frame on a Pixel 8a.

• Architecture: MiniFASNetV2 (depthwise-separable CNN) — pure CNN.
• Weights: minivision-ai/Silent-Face-Anti-Spoofing (2.7_80x80_MiniFASNetV2) · Apache-2.0.
• Size: 1.85 MB.

A photographed face is correctly flagged as a replay/presentation attack. A live camera capture scores "live". Portrait: Unsplash (free license).

I/O

• Input: [1, 3, 80, 80] NCHW, BGR, x/255 — a face crop (~2.7× the face box, centered).
• Output: [1, 3] softmax — class 1 = live/real, classes 0 & 2 = spoof (print / replay). Live score = output[1]; argmax == 1 → live.

GPU conversion

MiniFASNetV2 is a pure CNN → fully GPU-compatible (168/168 nodes on the delegate, 1 partition; device corr 1.0, ~5 ms) with zero patches (PReLU lowers to GPU-clean relu ops). CPU-exact vs PyTorch (corr 1.0).

Minimal usage

Kotlin (Android, LiteRT CompiledModel GPU)

val options = CompiledModel.Options(Accelerator.GPU)
val model = CompiledModel.create(context.assets, "silentface.tflite", options, null)
val inBufs = model.createInputBuffers()
val outBufs = model.createOutputBuffers()

inBufs[0].writeFloat(faceCropNCHW)       // [1,3,80,80] BGR, x/255
model.run(inBufs, outBufs)
val p = outBufs[0].readFloat()           // [3] softmax; live = p[1], spoof if argmax != 1

Python (LiteRT / ai-edge-litert)

import numpy as np
from ai_edge_litert.interpreter import Interpreter

it = Interpreter(model_path="silentface.tflite"); it.allocate_tensors()
inp, out = it.get_input_details(), it.get_output_details()
it.set_tensor(inp[0]["index"], x)        # [1,3,80,80] float32, BGR, x/255
it.invoke()
p = it.get_tensor(out[0]["index"])[0]    # [3]; live = p[1]

Note

The original repo ensembles two MiniFASNet models (crops at scale 2.7 and 4.0) with a face detector; this ships the primary MiniFASNetV2 (2.7). For best accuracy, feed a detected face crop and optionally ensemble the second model.

License

Apache-2.0 (Silent-Face-Anti-Spoofing / minivision-ai).