ProvenanceDeepfakeDetector<T>

Models & Types · AiDotNet.Safety.Image

Detects deepfake/AI-generated images by analyzing provenance signals: compression artifacts, statistical fingerprints, and embedded watermark traces.

Inheritance: ModelBase<T, Vector<T>, Vector<T>> → SafetyModuleBase<T> → ImageSafetyModuleBase<T> → DeepfakeDetectorBase<T> → ProvenanceDeepfakeDetector<T>

For Beginners

Real photos carry invisible “fingerprints” from the camera that took them — specific noise patterns, compression artifacts, and color processing signatures. AI images don’t have these, or they have different ones. This module checks for the presence or absence of these fingerprints to determine if an image is AI-generated.

How It Works

AI-generated images often lack natural camera processing artifacts (JPEG quantization patterns, sensor noise, optical aberrations) or contain telltale signs of specific generators (GAN checkerboard artifacts, diffusion model smoothing). This detector analyzes these provenance signals without requiring frequency domain analysis.

Detection signals:

  1. JPEG artifact analysis — real photos have characteristic quantization patterns
  2. Noise floor consistency — cameras produce consistent sensor noise
  3. Color channel correlation — natural images have specific cross-channel statistics
  4. Local Binary Pattern (LBP) texture analysis — AI textures differ from natural

References:

  • C2P-CLIP: Content provenance detection (2024, arxiv:2408.09647)
  • SynthID-Image: Internet-scale watermarking for provenance (DeepMind, 2025, arxiv:2510.09263)
  • Only 38% of AI generators have adequate watermarking (2025, arxiv:2503.18156)
  • AI-generated media detection survey (2025, arxiv:2502.05240)

Constructors

Constructor Summary
ProvenanceDeepfakeDetector(double threshold) Initializes a new provenance-based deepfake detector.

Properties

Property Summary
string ModuleName { get; }

Methods

Method Summary
IReadOnlyList<SafetyFinding> EvaluateImage(Tensor<T> image)
double GetDeepfakeScore(Tensor<T> image)