Skip to content

Latest commit

 

History

History
196 lines (141 loc) · 6.73 KB

README.md

File metadata and controls

196 lines (141 loc) · 6.73 KB

AI Image Detection Benchmark

A toolkit for benchmarking AI image detection APIs against various types of AI-generated images.

Overview

This project creates a benchmark dataset of real and AI-generated images, then evaluates multiple commercial AI detection APIs to measure their effectiveness. The goal is to assess the reliability of current detection tools across different generation methods, image categories, and adversarial perturbations.

Setup

1. Install Dependencies

uv pip install -r requirements.txt

2. Configure Environment

Copy the example environment file and fill in your API keys:

cp .env.example .env

Edit .env with your API credentials:

  • Image generation: Grok (xAI), FAL, Gemini, Replicate
  • Detection APIs: TruthScan, AIorNot, Resemble, SightEngine, etc.
  • LLM APIs: OpenAI (for image validation)
  • Storage: S3 bucket name and region

3. Database

The SQLite database (data/images.db) is auto-initialized on first run.

Architecture

Pipeline

Real Images (base)
  → validator.py                Validate base images with LLM
  → prompt_generator.py         Generate T2I prompts from base images
  → generator.py                Image editing (modify real images)
  → text_to_image.py            Text-to-image generation from prompts
  → adversarial.py              Adversarial perturbations on generated images
  → detector.py                 Run images through detection APIs
  → report.py                   Detection rate reports, heatmaps, ROC, AUC
  → report_adversarial_stats.py Adversarial effectiveness stats
  → report_deepfake_eval.py     Deepfake-Eval accuracy report

Database Schema

  • images - Base/real images with metadata, category, group, validation status
  • image_refs - S3 paths for both base and generated images
  • generated_images - Generated image records (generator, model, base_image_id)
  • detection_results - Detection API results (detector, ai_score, raw_response)

Categories

Category Description Generation Type
car_add_damage Cars with AI-added damage Image editing
delivery_proof Porches with AI-added packages Image editing
product Products with AI-added defects Image editing
receipts Receipts with AI-modified amounts Image editing
construction_site Construction site images Text-to-image
food Food images Text-to-image
getty_editorial Editorial/news photos Text-to-image
time_top_100 Iconic photographs Text-to-image

Groups

Images are organized into groups for different benchmark runs:

  • Group 2: ~1,563 base images (varying per category)
  • Group 3: 120 base images (15 per category, balanced)

Detectors

Detector API Score Interpretation
aiornot AI or Not ai.confidence (0-1)
truthscan TruthScan 0.5 +/- confidence/200 based on label
resemble Resemble AI image_metrics.score (0-1, higher=AI)
sightengine SightEngine type.ai_generated
realitydefender Reality Defender resultsSummary.metadata.finalScore/100
winston Winston AI ai_probability (0-1)
illuminarty Illuminarty data.probability (0-1)
inaza Inaza overall_risk_score (0-1)

Generators

Image Editing

  • grok-imagine-image-beta - Grok image editing
  • gpt-image-1.5 - GPT image editing
  • gemini-3-pro-image-preview - Gemini image editing

Text-to-Image

  • grok-imagine-image-beta - Grok T2I
  • gpt-image-1.5 - GPT T2I
  • gemini-3-pro-image-preview - Gemini T2I
  • qwen-image-2512 - Qwen T2I
  • seedream-v4.5-1k - Seedream T2I

Usage

1. Validate Base Images

# Validate base images using LLM (reject bad/unusable images)
python validator.py pixabay car_add_damage --limit 50
python validator.py pixabay food --batch-size 4

2. Generate Prompts (for T2I categories)

# Generate T2I prompts from base images using Gemini
python prompt_generator.py batch --category food --group 3
python prompt_generator.py batch --category construction_site --group 3 --limit 20

3. Generate Images

# Image editing (modify real images)
python generator.py --category car_add_damage --generator grok --group 3

# Text-to-image (from generated prompts)
python text_to_image.py --category food --generator openai-t2i --group 3

# Adversarial perturbations (on existing generated images)
python adversarial.py batch --source-generator grok --group 3

4. Run Detection

# Detect generated images with a specific detector
python detector.py -d resemble -t generated --group 3

# Detect base images (for false positive rates)
python detector.py -d aiornot -t base --group 3

# Limit and randomize
python detector.py -d resemble -t generated --limit 100 --random

5. View results

# Basic detection rate report
python report.py --group 3

# With adversarial comparison
python report.py --group 3 --adversarial

# With adversarial AUC and paired analysis
python report.py --group 3 --adversarial-auc --adversarial-paired

# Heatmaps and ROC curves
python report.py --group 3 --detection-heatmap --effect-size-heatmap --aggregate-roc

# LaTeX tables for paper
python report.py --group 3 --latex

Key Files

File Purpose
config.py Configuration and environment loading
storage.py Database and S3 operations
detector.py Detection API implementations
report.py Reporting and visualization
generator.py Image editing generation
text_to_image.py Text-to-image generation
adversarial.py Adversarial perturbation pipeline
prompt_generator.py LLM prompt generation for T2I
validator.py LLM-based image validation
check_c2pa.py C2PA metadata checker
image_utils.py Shared image utilities

Detection Score Interpretation

All detectors normalize scores to 0-1 where:

  • 0.0 = Definitely real
  • 0.5 = Uncertain
  • 1.0 = Definitely AI-generated

Default detection threshold is 0.5.