human_security 0.0.1

On-device face liveness verification with ML Kit active challenges, TFLite anti-spoofing, and optional face identity matching.

human_security

On-device face liveness verification for Flutter. Combines active ML Kit challenges (blink, smile, head turns) with passive TensorFlow Lite anti-spoofing, and optional face identity matching against a reference photo.

All inference runs locally on the device — no server calls are required for liveness or face matching (except when loading a reference image from a URL).

Features

Active liveness (Google ML Kit)

• Randomized challenge sequence (blink, smile, turn head left, turn head right)
• Face positioning guidance with oval overlay
• Real-time instructions and progress indicator
• Configurable number of challenges (challengeCount, default 2)
• Challenge timeouts and warmup periods to reduce false negatives
• Front/back camera flip button (when multiple cameras are available)

Passive anti-spoofing (TensorFlow Lite)

• Bundled anti_spoof.tflite model — no download step required
• Runs during challenge and evaluation phases with frame sampling
• JPEG capture fallback when stream-based passive scores are unavailable
• Combined active + passive scoring (60% active, 40% passive by default)
• Spoof detection for photos, screens, and static replays
• Can be disabled (enablePassiveLiveness: false) for active-only mode

Face identity matching (optional)

• Compare live capture against a reference image using face_embedding.tflite
• Reference sources: local file, in-memory bytes, or HTTP/HTTPS URL
• Cosine similarity with configurable threshold (matchThreshold, default 0.75)
• Returns a unified FaceMatchReport with liveness + match results

UI widget

• Drop-in FaceLivenessDetector widget with camera preview
• Customizable colors, instruction text style, and progress bar
• Success callback returns captured JPEG File and optional FaceMatchReport
• Failure callback with human-readable reason
• Built-in error view with retry

Lower-level APIs (exported)

Export	Purpose
FaceLivenessDetector	Main camera + liveness UI widget
FaceMatchService	Standalone face embedding comparison
ReferenceImageSource	File / bytes / URL reference input
FaceMatchReport	Combined verification result
LivenessChallenge	Challenge action enum
LivenessState / LivenessStatus	Liveness flow state
VerificationResult	Active + passive score breakdown
ScoreEvaluator	Threshold and scoring configuration
EmbeddingMath	Cosine similarity helpers

Platform support

Platform	Support
Android	✅ API 24+
iOS	✅ iOS 15.5+
Web	❌
Desktop	❌

Getting started

1. Add dependency

dependencies:
  human_security: ^0.0.1

2. Android configuration

android/app/src/main/AndroidManifest.xml — add permissions inside <manifest>:

<uses-permission android:name="android.permission.CAMERA" />
<uses-permission android:name="android.permission.INTERNET" />
<uses-feature android:name="android.hardware.camera" android:required="false" />
<uses-feature android:name="android.hardware.camera.front" android:required="false" />

android/app/build.gradle.kts (or build.gradle) — minimum SDK and TFLite settings:

android {
    defaultConfig {
        minSdk = 24
    }

    androidResources {
        // Prevent Gradle from compressing bundled .tflite assets
        noCompress += "tflite"
    }
}

// Pin TensorFlow Lite versions used by tflite_flutter (recommended)
configurations.all {
    resolutionStrategy {
        force("org.tensorflow:tensorflow-lite:2.16.1")
        force("org.tensorflow:tensorflow-lite-gpu:2.16.1")
        force("org.tensorflow:tensorflow-lite-api:2.16.1")
    }
}

Runtime: request camera permission before showing the liveness UI (e.g. with permission_handler).

3. iOS configuration

Minimum deployment target: iOS 15.5 (required by google_mlkit_face_detection).

ios/Podfile:

platform :ios, '15.5'

post_install do |installer|
  installer.pods_project.targets.each do |target|
    flutter_additional_ios_build_settings(target)
    target.build_configurations.each do |config|
      config.build_settings['IPHONEOS_DEPLOYMENT_TARGET'] = '15.5'
    end
  end
end

ios/Runner/Info.plist — privacy usage descriptions:

<key>NSCameraUsageDescription</key>
<string>Camera access is required for face liveness verification.</string>
<key>NSPhotoLibraryUsageDescription</key>
<string>Photo library access is required to pick a reference face image.</string>

If you load reference images from arbitrary URLs, ensure App Transport Security allows those hosts or use HTTPS.

4. Bundled models

The package ships two TFLite models under assets/models/:

Asset	Role
anti_spoof.tflite	Passive liveness / anti-spoof scoring
face_embedding.tflite	Face embedding extraction for identity match

These are declared in the package pubspec.yaml and loaded via default asset paths — no extra setup in your app.

Usage

Liveness only

import 'dart:io';

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

class LivenessPage extends StatelessWidget {
  const LivenessPage({super.key});

  @override
  Widget build(BuildContext context) {
    return Scaffold(
      body: FaceLivenessDetector(
        onSuccess: (File image, FaceMatchReport? report) {
          Navigator.pop(context);
          // image = captured live selfie
          // report?.livenessResult has active/passive/combined scores
        },
        onFailure: (String reason) {
          Navigator.pop(context);
          // Show reason to user
        },
      ),
    );
  }
}

Liveness + face match against a reference photo

FaceLivenessDetector(
  referenceImage: ReferenceImageSource.file(referenceFile),
  // or: ReferenceImageSource.url('https://example.com/id-photo.jpg')
  // or: ReferenceImageSource.bytes(imageBytes)
  matchThreshold: 0.75,
  onSuccess: (File image, FaceMatchReport? report) {
    if (report?.isVerified ?? false) {
      // Same person + live face
    }
  },
  onFailure: (String reason) {
    // Liveness failed or face did not match reference
  },
)

Active-only mode (no TFLite anti-spoof)

FaceLivenessDetector(
  enablePassiveLiveness: false,
  onSuccess: (image, report) { /* ... */ },
  onFailure: (reason) { /* ... */ },
)

Customization

FaceLivenessDetector(
  challengeCount: 3,
  matchThreshold: 0.80,
  backgroundColor: Colors.black,
  overlayColor: const Color(0xCC000000),
  progressColor: Colors.greenAccent,
  showCameraFlipButton: true,
  instructionTextStyle: const TextStyle(
    color: Colors.white,
    fontSize: 20,
    fontWeight: FontWeight.w600,
  ),
  onSuccess: (image, report) { /* ... */ },
  onFailure: (reason) { /* ... */ },
)

Verification flow

flowchart TD
    A[Initialize camera + models] --> B[Position face in oval]
    B --> C[Active challenges]
    C --> D[Passive anti-spoof collection]
    D --> E{Scores pass?}
    E -->|No| F[onFailure]
    E -->|Yes| G[Capture photo]
    G --> H{Reference image set?}
    H -->|No| I[onSuccess with liveness report]
    H -->|Yes| J[Compare embeddings]
    J --> K{Match + live?}
    K -->|Yes| I
    K -->|No| F

Scoring defaults

ScoreEvaluator defaults:

Setting	Default	Description
passiveThreshold	0.5	Minimum average passive score to count as live
combinedThreshold	0.75	Minimum combined active + passive score
activeWeight	0.6	Weight of active challenge completion
passiveWeight	0.4	Weight of passive anti-spoof score
strongSpoofThreshold	0.35	Frames below this trigger fast spoof failure

Face match uses matchThreshold on cosine similarity (default 0.75).

Example app

See the example/ directory for a full demo with:

• Full liveness (active + passive)
• Active-only mode
• Optional reference image from gallery or URL
• Result screen with scores and side-by-side comparison

cd example
flutter pub get
cd ios && pod install && cd ..
flutter run

Dependencies

• camera — live camera preview and capture
• google_mlkit_face_detection — face landmarks and active challenges
• tflite_flutter — on-device TFLite inference
• image — JPEG decode for passive fallback
• http — optional reference image download

Additional information

• Repository: github.com/ShithinCherathuparambil/human_security
• Issues: GitHub Issues
• License: MIT — see LICENSE