liveness_edge_flutter 0.3.0

Private, offline face-liveness verification for Flutter using MediaPipe and ONNX Runtime.

Liveness Edge Flutter

[Liveness Edge Flutter: on-device processing, randomized active challenges, passive anti-spoofing, and face continuity]

Private, offline face-liveness verification for Flutter. Camera frames, face landmarks, active challenges, and passive anti-spoofing are processed on the device; the package does not make network requests.

Features

• Ready-to-use camera screen with face-position guidance.
• Three or four active challenges randomly selected from blink, one randomly directed head turn, smile, and open mouth.
• Passive presentation-attack score using MiniFASNetV2.
• Same-face continuity checks using normalized facial landmarks.
• Native MediaPipe Face Landmarker and ONNX Runtime processing.
• Final verified capture returned as JPEG bytes.
• Localizable messages and customizable guidance, supporting text, and retry UI.
• Android and iOS support with bundled models; no backend required.

Platform support

Platform	Minimum version	Status
Android	API 24	Supported
iOS	15.1	Supported

The host device needs a front-facing camera. Desktop and web are not supported.

Installation

Add the package to your application:

flutter pub add liveness_edge_flutter

Alternatively, add it directly to pubspec.yaml:

dependencies:
  liveness_edge_flutter: ^0.3.0

Android setup

Add camera permission to android/app/src/main/AndroidManifest.xml:

<uses-permission android:name="android.permission.CAMERA" />

Ensure the application has minSdk 24 or newer.

iOS setup

Add a camera usage description to ios/Runner/Info.plist:

<key>NSCameraUsageDescription</key>
<string>The camera is used to verify that you are present.</string>

Set the application deployment target to iOS 15.1 or newer.

Usage

Import the package and push LivenessEdgeScreen as a full-screen route:

import 'package:flutter/material.dart';
import 'package:liveness_edge_flutter/liveness_edge_flutter.dart';

Future<LivenessResult?> captureLiveness(BuildContext context) async {
  LivenessResult? verifiedResult;

  await Navigator.of(context).push<void>(
    MaterialPageRoute(
      builder: (routeContext) => LivenessEdgeScreen(
        configuration: const LivenessConfiguration(
          validations: {
            LivenessValidation.blink,
            LivenessValidation.headTurn,
            LivenessValidation.smile,
            LivenessValidation.openMouth,
            LivenessValidation.passiveAntiSpoof,
          },
          timeout: Duration(minutes: 2),
          maxFrames: 3000,
          passiveAntiSpoofSensitivity: PassiveAntiSpoofSensitivity.high,
          faceIdentitySensitivity: FaceIdentitySensitivity.strict,
        ),
        onSuccess: (result) {
          verifiedResult = result;
          Navigator.of(routeContext).pop();
        },
        onFailed: (result) {
          debugPrint('Liveness failed: ${result.instruction}');
        },
        onCancel: () => Navigator.of(routeContext).pop(),
      ),
    ),
  );

  return verifiedResult;
}

On success, result.imageBytes contains the verified JPEG and result.passiveScore contains the median passive anti-spoof score. The package does not upload or persist either value.

All user-facing text defaults to English. Override any message through the configuration to localize the flow or use your own product copy:

const LivenessConfiguration(
  messages: LivenessMessages(
    moveRight: 'Geser wajah ke kanan',
    moveLeft: 'Geser wajah ke kiri',
    blink: 'Kedipkan kedua mata sekali.',
    turnLeft: 'Hadapkan kepala sedikit ke kiri.',
    turnRight: 'Hadapkan kepala sedikit ke kanan.',
    smile: 'Tersenyumlah.',
    openMouth: 'Buka mulut.',
    returnToNeutral: 'Kembali ke ekspresi netral.',
    tryAgain: 'Coba lagi',
  ),
);

Colors, the camera-guide shape, an optional header, text, and the retry button can use the host application's design system. Text styles are merged with the themed defaults, while builders replace their corresponding built-in content:

LivenessEdgeScreen(
  theme: const LivenessEdgeTheme(
    backgroundColor: Colors.black,
    foregroundColor: Colors.white,
    primaryColor: Colors.orange,
    successColor: Colors.green,
    inactiveRingColor: Colors.white24,
    cameraShape: LivenessCameraShape.roundedRectangle,
    cameraBorderRadius: 32,
    guidelineTextStyle: TextStyle(
      fontSize: 23,
      fontWeight: FontWeight.w700,
    ),
    supportingTextStyle: TextStyle(fontSize: 14),
  ),
  headerBuilder: (context, state) => const CompanyLogo(),
  retryButtonBuilder: (context, label, onPressed) => OutlinedButton.icon(
    onPressed: onPressed,
    icon: const Icon(Icons.replay_rounded),
    label: Text(label),
  ),
);

cameraShape supports oval (the default), circle, and roundedRectangle. cameraBorderRadius applies to rounded rectangles. The header receives a read-only LivenessEdgeViewState containing the current status, instruction, supporting text, progress, camera readiness, and error state. The legacy guidelineTextStyle and supportingTextStyle screen properties remain available for backward compatibility but are deprecated.

Configuration

Property	Default	Description
validations	All validations	Enabled checks: blink, headTurn, smile, openMouth, and/or passiveAntiSpoof. headTurn randomly chooses either a left or right turn for the session.
minimumActiveChallenges	3	Minimum active actions selected per session, capped by the available pool.
maximumActiveChallenges	4	Maximum active actions selected per session, capped by the available pool.
timeout	2 minutes	Maximum session duration.
maxFrames	3000	Safety cap for analyzed frames. Wall-clock duration is controlled by timeout.
passiveAntiSpoofSensitivity	balanced	Anti-spoof preset: low (0.40), balanced (0.50), high (0.60), strict (0.70), or withValue(...) for a custom value from 0 to 1.
faceIdentitySensitivity	balanced	Face replacement preset: low (0.16), balanced (0.10), high (0.06), strict (0.035), or withValue(...) for a custom value greater than 0.
messages	English LivenessMessages	User-facing challenge instructions and screen labels. Override individual values to localize the flow.

Treat the default threshold as a starting point. Calibrate it using genuine users, target devices, lighting conditions, and representative attacks. At least one validation must be enabled. A session securely shuffles its active action pool and selects three or four actions without duplicates by default. Because headTurn contributes one randomly directed action, the default pool contains four active actions. If fewer active validations are enabled, the configured minimum and maximum are capped to the number available. A passive-only session is also supported by enabling only passiveAntiSpoof. Only the current instruction is displayed. Each gesture must be observed across multiple frames and return to its neutral pose before the next action starts. Disabled active challenges are skipped, so their instructions are not shown. Passive sensitivity is ignored when passiveAntiSpoof is disabled.

After the configured checks pass, the package waits for three stable, frontal frames with both eyes fully open before producing the verified JPEG.

The package keeps a frozen, normalized, on-device landmark descriptor for the first aligned face. If a different frontal face is detected in two consecutive frames, the active challenge returns to alignment. A session cannot pass while an identity mismatch is pending. If no valid single face is detected during an active challenge for 1.2 seconds, the flow also returns to alignment. Before a challenge starts, the grace period remains two seconds.

Callback behavior

• onSuccess runs once after both checks pass. It does not close the route; the host application controls navigation.
• onFailed runs once when an attempt fails because it times out, exceeds the frame limit, or does not pass the passive anti-spoof check. Retrying starts a new attempt and may invoke it again. It does not close the route.
• onCancel runs when the close button is pressed. When omitted, the screen attempts to pop its route automatically.
• A failed or timed-out session shows a retry action. Camera or model errors do not invoke onFailed, because they do not produce a final liveness result.

Result fields

Property	Availability	Description
status	Always	Current or terminal LivenessStatus. A terminal result is either passed or failed.
passed	Always	Convenience getter that is true only when status is passed.
instruction	Always	Localized guidance or the terminal result message.
framesProcessed	Always	Number of analyzed frames in the current attempt.
imageBytes	Success only	Mirrored, portrait-normalized JPEG cropped to the face-guide aspect ratio.
passiveScore	Terminal result when passive anti-spoofing is enabled	Median of the recent on-device live-score samples. Higher values are considered more likely to be live.

How it works

Package integration flow

The android/ and ios/ directories at the package root contain the native plugin implementations that are included in every host application. The directories with the same names under example/ are only the Android and iOS hosts used to run the sample application.

Host application (or example/lib/main.dart)
  -> package public API (lib/liveness_edge_flutter.dart)
  -> LivenessEdgeScreen and Dart challenge state machine
  -> Flutter MethodChannel
       -> Android: android/.../LivenessEdgeFlutterPlugin.kt
       -> iOS:     ios/Classes/LivenessEdgeFlutterPlugin.swift
  -> MediaPipe Face Landmarker + ONNX Runtime
  -> LivenessResult returned to the host application

The relevant package layout is:

liveness_edge_flutter/
├── lib/                 public API and Dart implementation
├── android/             native Android plugin implementation
├── ios/                 native iOS plugin implementation
└── example/
    ├── lib/             sample Flutter application
    ├── android/         Android host for the sample
    └── ios/             iOS host for the sample

Frame processing flow

Flutter camera frame
  -> upright, downsampled BGR frame
  -> MediaPipe Face Landmarker (bounds, blink, expression, and head yaw)
  -> MiniFASNetV2 through ONNX Runtime (passive live score)
  -> challenge state machine (align, 3-4 randomized actions, neutral return,
     same-face continuity, stable final capture)
  -> LivenessResult

Models are bundled with the native Android and iOS implementations. Processing stays in application memory unless the host application stores or transmits the returned image.

Security and privacy

Client-side liveness is not tamper-proof. Do not use result.passed as the sole authorization for banking, e-KYC, account recovery, or another high-risk server transaction. For those flows, combine local inference with a server-issued nonce, platform attestation, replay protection, and a trusted server-side decision.

The passive model primarily targets printed-photo and screen-replay attacks. It has not been validated by this project against masks, 3D attacks, camera injection, rooted devices, or modified applications. The host application is responsible for consent, retention, logging, screenshots, backups, and privacy disclosures.

Review the licenses and suitability of MediaPipe, ONNX Runtime, and the bundled MiniFASNetV2 model before commercial distribution.

Troubleshooting

• Camera permission denied: request permission through the normal platform flow and ensure the manifest or plist entry above is present.
• Front camera unavailable: the capture screen reports an error and allows retrying; use a physical device with a front-facing camera.
• Frequent passive failures: test representative lighting and devices before lowering passiveAntiSpoofSensitivity; lower values increase spoof acceptance risk.
• Simulator or emulator: camera and native inference behavior can differ from physical hardware, so release validation should use real devices.

Example and development

See example/lib/main.dart for a runnable integration.

flutter analyze
flutter test
cd example
flutter run

Bug reports and feature requests are welcome in the project issue tracker.

Support

If this package is useful to you, you can support its development through Buy Me a Coffee.