40% Poids de Détection

Détection ML

Modèle Hugging Face Transformer

Notre détection ML utilise des modèles Transformer entraînés sur des millions d'images.

92%
Précision Moyenne
40%
Poids de Détection
ML Detection Illustration

Comment ça Fonctionne

Le réseau neuronal analyse les motifs visuels, textures et artefacts subtils invisibles à l'œil humain mais caractéristiques de la génération IA. Il examine les caractéristiques au niveau des pixels, les distributions de couleurs et les motifs structurels appris à partir d'images réelles et générées par IA.

1

Image Input

Upload any image file

2

Preprocessing

Resize and normalize

3

Analysis

Transformer model inference

4

Result

AI probability score

Technical Specifications

Model Details

  • Model ViT-based Classifier
  • Source Hugging Face
  • Training Data 1M+ images
  • Processing CPU-optimized

Detection Capabilities

  • Stable Diffusion variants
  • Midjourney outputs
  • DALL-E generations
  • Other diffusion models

Frequently Asked Questions

What is Machine Learning detection for AI images?

ML detection uses trained neural networks (specifically Vision Transformers) to analyze images and identify patterns characteristic of AI-generated content, such as those from Stable Diffusion, DALL-E, or Midjourney.

Why is ML detection weighted at 40%?

ML detection is our most accurate single method, trained on millions of images. It achieves 92-98% accuracy on direct AI outputs, making it the primary signal in our ensemble detection system.

What AI generators can ML detection identify?

Our ML model detects images from Stable Diffusion (all versions), DALL-E 2 & 3, Midjourney, Adobe Firefly, Leonardo.ai, and most diffusion-based generators.

How does the Vision Transformer work?

Vision Transformers (ViT) divide images into patches and learn attention patterns between them. They can identify subtle correlations that differ between AI-generated and real photographs.

Does image compression affect ML detection?

Moderate JPEG compression (quality 60-100) has minimal impact. Heavy compression or multiple re-compressions can reduce accuracy, which is why we use ensemble methods.

What image formats are supported?

We support JPEG, PNG, WebP, BMP, and TIFF formats. All images are preprocessed to 224x224 pixels while preserving aspect ratio for optimal model inference.

How fast is the ML detection?

Our CPU-optimized model processes images in under 500ms. GPU acceleration can reduce this to under 50ms for batch processing.

Can ML detection identify edited photos?

ML detection focuses on fully AI-generated images. For AI-enhanced or partially edited photos, other methods like clone detection work better.

Is the model updated for new AI generators?

Yes, we regularly retrain our models to include outputs from the latest AI image generators. The current model is trained on images from 2024-2026 generators.

What is the false positive rate?

Our ML model has a false positive rate under 2% for standard photographs. Heavily filtered or stylized photos may occasionally trigger false positives, which ensemble methods help mitigate.

Méthodes associées

Analyse PRNU

Photo Response Non-Uniformity (PRNU) detects unique camera sensor fingerprints from manufacturing imperfections. AI images cannot replicate these authentic sensor signatures.

Analyse Fréquentielle

L'analyse du domaine fréquentiel examine la distribution des composantes haute et basse fréquence d'une image. Les images générées par IA manquent généralement du bruit naturel haute fréquence présent dans les vraies photographies.

Analyse de gradient

Analyzes edge patterns and texture characteristics using Sobel, Canny, and Laplacian operators. AI images often have unnaturally smooth or uniform gradients.

Motif de bruit

Les vraies photographies contiennent des motifs de bruit uniques des capteurs de caméra qui varient à travers l'image. Les images générées par IA ont une distribution de bruit anormalement uniforme.

Analyse des métadonnées

Les métadonnées d'image contiennent des indices précieux sur son origine. Nous analysons les données EXIF, les signatures logicielles et autres informations intégrées pour identifier les outils de génération IA.

Empreinte GAN

Détecte les artefacts spécifiques aux GAN comme les motifs en damier et le banding couleur.

Analyse de texture

Analyse Local Binary Pattern pour les anomalies de texture dans les images IA.

Détection Anatomique

Les générateurs d'images IA créent souvent des erreurs anatomiques que les humains reconnaissent immédiatement comme fausses. Nous utilisons la vision par ordinateur pour détecter ces erreurs révélatrices.

Vérification C2PA

C2PA (Coalition for Content Provenance and Authenticity) est un standard industriel pour suivre l'origine et l'historique du contenu numérique via des signatures cryptographiques.

Semantic Inconsistency Detection

Detects logical inconsistencies like incorrect shadows, impossible perspectives, distorted reflections, and violations of physical laws that AI often produces.

Human Biometric Analysis

Uses MediaPipe to analyze human anatomy for incorrect finger counts, asymmetric eyes, unnatural skin texture, and other anatomical anomalies common in AI-generated faces.

Lighting Physics Validation

Validates light source consistency, shadow direction physics, specular highlight accuracy, and color temperature uniformity across the image.

Compression Artifact Analysis

Analyzes JPEG compression artifacts to estimate quality levels and detect re-compression patterns that indicate image manipulation or AI generation.

Edge Sharpness Analysis

Analyzes sharpness distribution across the image and validates depth-of-field consistency. AI often produces unnaturally uniform sharpness.

Statistical Pattern Analysis

Analyzes statistical properties including Shannon entropy, histogram patterns, and Benford's Law compliance to detect synthetic image characteristics.

Chromatic Aberration Analysis

Detects the absence of chromatic aberration (color fringing) that real camera lenses produce. AI images lack these optical artifacts.

Micro-Texture Analysis

Analyzes microscopic texture patterns for repetition, uniformity, and unnatural randomness that AI generators often exhibit.

Color Palette Analysis

Analyzes color distribution including saturation levels, color diversity, and white balance consistency. AI images often have oversaturated colors.

Vérifier Votre Image

Toutes les méthodes sont combinées en utilisant un score pondéré pour produire un verdict final avec niveau de confiance.

Essayer Maintenant