spoof-detection-package

On-device face anti-spoofing for Expo and React Native. Give it a photo uri, it tells you whether the face is real or a presentation attack — printed photo, passport photo, poster, or a phone/tablet/laptop screen.

Everything runs natively on-device (ML Kit / Apple Vision for face detection, MiniFASNetV2 for texture analysis — converted from the official Silent-Face-Anti-Spoofing weights and bundled in the package, ~1.9 MB). No network calls, no photo ever leaves the device, no JS dependencies.

Install

npx expo install spoof-detection-package
# bare React Native: also needs expo-modules-core set up (npx install-expo-modules)

Native module — needs a dev build / prebuild, does not work in Expo Go.

Plays nice with other TFLite packages

Safe to install alongside react-native-fast-tflite (or alone):

• Android uses Google's LiteRT artifact (com.google.ai.edge.litert:litert), the same family fast-tflite v3 uses — Gradle dedupes to a single copy, no duplicate-class or duplicate-.so errors.
• iOS declares TensorFlowLiteSwift >= 2.14, < 3.0 as a range, so CocoaPods can settle on whatever exact TensorFlowLiteC version other pods pin (e.g. fast-tflite 3.x pins 2.17.0) instead of conflicting.

Usage

import { detectSpoof, isSpoof } from 'spoof-detection-package';

const result = await detectSpoof(photo.uri);
// {
//   live: true,          // safe to proceed
//   spoof: false,        // presentation attack detected
//   realScore: 0.94,     // P(real face) from the model, 0-1
//   reason: null,        // 'no_face' | 'multiple_faces' | 'face_too_small' | 'texture' | null
// }

if (result.spoof) {
  // show "Spoof detected" popup
} else if (!result.live) {
  // quality problem: result.reason === 'no_face' or 'face_too_small'
} else {
  // real face — continue with recognition / attendance
}

// or the one-liner:
if (await isSpoof(photo.uri)) return blockScan();

Options

await detectSpoof(uri, {
  threshold: 0.7,     // min P(real) to accept; raise = stricter
  minFaceRatio: 0.18, // face must be >= 18% of the photo's short side
  cropScale: 2.7,     // context crop around the face (model default)
});

How it decides

1. Face detection (ML Kit on Android, Apple Vision on iOS)
   • no face -> { live: false, spoof: false, reason: 'no_face' }
   • 2+ faces -> { spoof: true, reason: 'multiple_faces' }
   • tiny face -> { live: false, reason: 'face_too_small' } (ask the user to come closer — also blocks passport photos held at a distance)
2. MiniFASNetV2 texture analysis on a wide 2.7x crop (sees paper edges, screen bezels, glare, moiré) -> realScore; below threshold -> { spoof: true, reason: 'texture' }

Tips

• Combine with an object detector (COCO-SSD "cell phone"/"tv"/"laptop") and a motion check across preview frames for defense in depth.
• Tune threshold with real data: log realScore for genuine users and known fakes, pick the value that separates them.