Image Detection

Image Detection

Manuscript analyzes forensic signals in image files to detect AI-generated content.

Detection Signals

EXIF Metadata Analysis

Real photographs contain EXIF data from the camera:

Metadata	Real Photo	AI-Generated
Camera make	Apple, Canon, Sony	Missing or generic
Camera model	iPhone 15, EOS R5	Missing
GPS coordinates	Present (often)	Never
Date/time	Consistent	Missing or fake
Exposure settings	Physical values	Missing

Weight: 0.25

Sensor Noise Pattern

Real cameras produce characteristic noise patterns. AI images lack authentic sensor noise.

Weight: 0.15

Color Distribution

Human photographs have natural color histograms. AI often produces artificially smooth distributions.

Weight: 0.20

Compression Artifacts

JPEG compression leaves specific patterns. AI images may have inconsistent compression artifacts.

Weight: 0.15

Edge Consistency

AI images often have unnatural edges, especially around fine details like hair, fingers, and text.

Weight: 0.15

Symmetry Detection

AI-generated faces often exhibit unnatural symmetry that real faces don’t have.

Weight: 0.10

Example Analysis

Real photograph:

File: vacation_photo.jpg
EXIF: Canon EOS R5, 1/250s, f/4, ISO 400
GPS: 40.7128° N, 74.0060° W
Date: 2024-12-25 14:30:00

Signals detected:

• Valid EXIF metadata ✓
• Authentic camera signature ✓
• Natural noise pattern ✓
• Consistent compression ✓

Verdict: Human (confidence: 0.89)

---

AI-generated image:

File: portrait.png
EXIF: None
GPS: None
Date: None

Signals detected:

• No EXIF metadata ✗
• No camera signature ✗
• Too clean (no sensor noise) ✗
• Symmetric face detected ✗

Verdict: AI (confidence: 0.85)

API Usage

Upload via Form

curl -X POST http://localhost:8080/verify \
  -F "image=@photo.jpg"

Base64 Encoded

curl -X POST http://localhost:8080/verify \
  -H "Content-Type: application/json" \
  -d '{
    "image_base64": "data:image/jpeg;base64,/9j/4AAQSkZJ..."
  }'

Response

{
  "id": "hm_xyz789",
  "verdict": "ai",
  "confidence": 0.76,
  "content_type": "image",
  "signals": {
    "metadata_score": 0.0,
    "color_distribution": 0.65,
    "edge_consistency": 0.42,
    "noise_pattern": 0.35,
    "compression_analysis": 0.58,
    "symmetry_score": 0.78
  },
  "processing_time_ms": 45
}

Current Benchmarks

Metric	Value	Notes
Accuracy	50.00%	Needs improvement
Precision	50.00%
Recall	100.00%	All AI images detected
F1 Score	66.67%

Known Issues

The current accuracy is limited because:

1. Web-sourced images often have EXIF metadata stripped
2. Downloaded photos lose camera signatures
3. Social media compression removes forensic markers

This is a known limitation. See Improvement Stories for planned enhancements.

Supported Formats

• JPEG (.jpg, .jpeg)
• PNG (.png)
• WebP (.webp)
• GIF (.gif) - first frame
• BMP (.bmp)
• TIFF (.tif, .tiff)

Best Practices

1. Original files: Detection works best on original camera files
2. Uncompressed: Avoid heavily compressed images
3. Full resolution: Don’t resize before analysis
4. Multiple samples: For important decisions, analyze multiple images

Planned Improvements

• DCT coefficient analysis for JPEG forensics
• Full EXIF parsing with validation
• PNG chunk analysis
• CLIP-based deep analysis (optional)