passive_liveness 0.1.1

Ultra-lightweight passive face liveness detection using native TFLite (Google Play Services on Android, TensorFlowLiteSwift on iOS) for Flutter.

0.1.1

• Expanded Flutter & Dart Compatibility: Widened environment constraints to support Dart SDK sdk: ">=3.0.0 <4.0.0" and Flutter flutter: ">=3.10.0".

0.1.0

• Low-Light Face Liveness Auto-Acceptance (lowLightThreshold): Added configurable low-light auto-acceptance to eliminate false spoof rejections and errors in dark/dimly-lit environments (such as night shifts or low-light work spaces).
  • lowLightThreshold (double?, default null): When specified (e.g. 70.0 or 0.30), frames captured with mean luminance below this threshold are automatically accepted as live (isReal: true, status: LivenessStatus.real), bypassing dark-induced neural logit suppression. Supports both 0..255 direct scale and normalized 0..1 scale.
  • Exposed Luminance Metrics: LivenessResult provides meanLuminance ($0.0 \dots 255.0$) and isLowLight for diagnostic logging and UI display.
• Upgraded Multi-Modal Anti-Spoofing Heuristics & Decision Fusion: Enhanced resilience against high-resolution OLED / Retina / MacBook screen replay attacks without increasing package size or native binary footprint.
  • 2D Flatness Check (is2DFlatSpoof): Evaluates Laplacian depth-of-field variance delta between center face crop ($\sigma^2_{\text{Face}}$) vs background region ($\sigma^2_{\text{Background}}$) to detect flat 2D focal planes ($\Delta < 0.08$).
  • Emissive Saturation Spikes (isEmissiveSaturationSpoof): Converted color space crops to HSV to analyze Saturation ($S$) channel variance ($\text{varSat} \ge 0.045$) relative to Hue, identifying additive RGB display backlight scatter.
  • Fast 2D Radix-2 FFT Moiré Analyzer (FftMoireAnalyzer): Pure Dart-only Cooley-Tukey FFT algorithm ($O(n \log n)$) to detect high-frequency sub-pixel display grid peaks ($\text{regularity} \ge 25.0$) beyond 60% Nyquist radius with zero binary size overhead.
  • Calibrated Multi-Factor Decision Fusion Engine: Re-calibrated decision override rules to reliably catch screen photo attacks (including low-brightness MacBook photos) while protecting genuine live face selfies under indoor lighting.
  • Enhanced Debug & Watermark Metrics: Exposed laplacianDelta, saturationVariance, and moireHighFreqRatio metrics in LivenessResult for detailed debugging and watermark logging.

0.0.6

• Direct CameraImage Liveness API Support (detectLivenessFromCameraImage): Added direct CameraImage evaluation support to PassiveLivenessDetector, eliminating redundant developer boilerplate for converting camera frames to LivenessImageBuffer.

0.0.5

• Native MethodChannel TFLite Architecture (~0MB Android APK Impact): Migrated model inference from flutter_litert Dart FFI to native platform channels (com.andikatp.passiveLiveness), reducing plugin binary size footprint from ~40MB down to almost 0MB.
  • Android: Uses Google Play Services TFLite runtime (play-services-tflite-java:16.5.0 & play-services-tflite-gpu:16.5.0), eliminating ~40MB APK/AAB size bloat by utilizing system-shared TFLite binaries with automatic GPU acceleration & CPU fallback.
  • iOS: Uses Apple-thinning optimized TensorFlowLiteSwift (~> 2.14) targeting iOS 12.0+.
• Zero-Dependency Plugin Architecture: Completely removed third-party flutter_litert dependency.
• Pure Dart Heuristic Processing: All anti-spoofing heuristic layers (FaceProximityGate, LbpHogAnalyzer, ColorSpaceAnalyzer, HighResScreenAnalyzer, and ImagePreprocessor) remain 100% in Dart for zero-overhead performance.
• Lower-end Device Compatibility (Low Light & High Sensor Noise): Re-enabled automatic brightness adjustment (enableContrastStretch) in the main inference pipeline to improve neural model accuracy on budget phones with dark environments.
• Chrominance Variance & Heuristic Decision Tuning: Calibrated ColorSpaceAnalyzer.maxVarianceThreshold to 140.0 and upgraded multi-factor decision fusion to reliably capture subtle OLED screen replays (e.g. spoof4) while eliminating false spoof rejections on genuine faces.
• Cleaner Diagnostics: Streamlined LivenessLogger to reflect Native Platform Channel engine initialization.
• Enhanced .gitignore: Added full coverage for Android and iOS build output directories (.gradle/, Pods/, DerivedData/, etc.).

0.0.4

• High-Resolution Screen Replay Anti-Spoofing (HighResScreenAnalyzer): Added 2D Laplacian frequency variance and patch focus depth dispersal ($\sigma^2_{\text{PatchLap}}$) analysis to catch high-resolution OLED/Retina/4K screen replay presentation attacks.
• Glasses Glare & Frame Edge False Positive Resolution: Added specular glare highlight masking ($\ge 245$ brightness) and multi-region upper-face HOG peak de-biasing (LbpHogAnalyzer & ColorSpaceAnalyzer) to prevent linear glasses frames and anti-reflective lens glare from triggering false spoof classifications on genuine users.
• Multi-Factor Liveness Decision Fusion Engine: Upgraded PassiveLivenessDetector decision engine to use multi-factor fusion scoring with adaptive neural model thresholding.
• Zero-Config Default Heuristic Engines: All anti-spoofing heuristic layers (enableTextureAnalysis, enableColorSpaceAnalysis, enableHighResScreenAnalysis) are now enabled by default (true).
• Streamlined API Parameter Signatures: Cleaned up redundant internal low-level flags across detectLivenessFromImageBytes, detectLivenessFromImageFile, detectLivenessFromBuffer, and LivenessFrameProcessor.processBufferFrame for a zero-boilerplate developer experience.
• Face Aspect Ratio & Proximity Gate: Added early pre-inference rejection gate (FaceProximityGate) to discard presentation attacks with small cropped photos ($<5%$ area), extreme close-ups ($>85%$ area), or distorted aspect ratios ($0.50 \dots 1.25$).
• Micro-Texture LBP / HOG Analysis Engine: Added LbpHogAnalyzer to evaluate $256 \times 256$ un-downscaled face crops for paper inkjet print halftone patterns (Local Binary Patterns) and screen grid alignments (Histogram of Oriented Gradients).
• YCbCr / YUV Color Space Analysis: Added ColorSpaceAnalyzer to calculate chrominance sub-sampling variance ($\sigma^2_{CbCr}$) directly from YUV camera streams to detect emissive RGB digital display screen replay attacks (iPad/tablet video replays).
• Adaptive Screen Flash Overlay: Added LivenessFlashController and AdaptiveScreenFlashOverlay UI components for active photometric stereo assist.
• New LivenessStatus Diagnostic Reason Codes: Introduced tooFar, tooClose, invalidAspectRatio, printSpoof, and screenReplaySpoof to give developers detailed status feedback.

0.0.3

• Android Sensor Coordinate Rotation & Accuracy Fixes: Fixed coordinate transformation for Android ML Kit portrait face bounding boxes across all sensor angles (0°, 90°, 180°, 270°), ensuring accurate upright crops and resolving false spoof detections on Android devices.
• Low-Light & Dark Environments Compensation: Enhanced anti-spoofing accuracy in dim lighting using adaptive power-law gamma expansion ($\gamma \approx 0.60 - 0.88$) and dynamic contrast stretching without corrupting facial texture liveness signals.
• Glasses Glare & Specular Reflection Resilience: Implemented Asymmetric Exponential Moving Average (EMA) score filtering ($\alpha=0.1$ for score drops, $\alpha=0.4$ for recovery) to prevent momentary specular lens reflections on glasses from triggering false spoof classifications.
• LiteRT Next CompiledModel Zero-Copy Acceleration: Integrated LiteRT Next CompiledModel support for zero-copy GPU/CPU hardware acceleration with automatic fallback to classic Interpreter.
• Flexible Model Tensor Layout Support: Added automatic shape inspection to support both NCHW ([1, 3, H, W]) and NHWC ([1, H, W, 3]) tensor models.
• Edge Pixel Replication (BORDER_REPLICATE): Improved image preprocessor boundary handling to eliminate pitch-black border artifacts on tight face crops.
• Pending Result & EMA Tracker Controls: Added LivenessResult.pending() factory constructor and resetEma() controls for handling motion transition frames cleanly.

0.0.2

• Fix Android ML Kit Bounding Box Coordinate System Mismatch: Fixed false spoof classification on Android caused by portrait bounding box offsets. Added isRotatedBoundingBox parameter and FaceBoundingBox.toRawBufferSpace() transformation for 0°, 90°, 180°, and 270° sensor angles.
• Fix Android 270° Front Camera Upside-Down Crop: Corrected case 270 coordinate mapping in ImagePreprocessor to ensure MiniFAS receives upright face crops.
• Offload Inference to Background Isolate (IsolateInterpreter): Prevented main UI isolate frame drops by running LiteRT inference asynchronously on a dedicated background Dart isolate.
• Add XNNPack ARM NEON SIMD Hardware Vectorization: Enabled XNNPackDelegate for 2x–4x CPU matrix multiplication speedup on mobile chipsets.
• Fast NV21 Plane Copy: Optimized _getNv21Bytes to replace 518,400 Dart loop iterations with native setRange array copies.
• Low-Light Adaptive Gamma Contrast Expansion: Replaced linear boost with adaptive power-law gamma curve ($\gamma \approx 0.60 - 0.88$) to expand 3D skin texture gradients in dim lighting.

0.0.1

• Initial release of passive_liveness.
• Ultra-lightweight passive face anti-spoofing engine powered by LiteRT (TensorFlow Lite) edge inference.
• Real-time zero-copy camera frame processing (LivenessImageBuffer) for high FPS video streams (iOS BGRA8888, Android NV21/YUV420).
• Support for static image file (File) and byte array (Uint8List) liveness detection (e.g. from takePicture() or image_picker).
• Closed-form reflect101 boundary padding and camera sensor rotation sampling (0°, 90°, 180°, 270°).
• Built-in low-light shadow-lift compensation for underexposed camera environments.