ERIC KIM

Oh yes—let’s build you a mini Image Forensics Lab you can run as a website and a CLI tool! 🎉

Written by

admin

in

I’ve packaged a complete, ready‑to‑run project that checks images for likely tampering using:

  • EXIF metadata inspection
  • ELA (Error Level Analysis)
  • Noise residual heatmap
  • Copy–move clone detection (ORB feature clustering)
  • JPEG “ghosts” curve + ghost variance map
  • A friendly heuristic suspicion score + verdict

💡 Reality check: no tool can prove an image is fake or real. This is a triage toolkit that gives you signals and a confidence score—great for first-pass vetting and learning.

⬇️ Grab the project

Download the project ZIP

Unzip it, and you’ll see:

image-forensics-lab/

├─ app.py                # Streamlit web app

├─ forensic.py           # Analysis routines (ELA, noise, copy–move, JPEG ghosts, scoring)

├─ cli.py                # Command-line tool

├─ requirements.txt      # Dependencies

├─ README.md             # Quick guide

└─ LICENSE               # MIT

🚀 Run the Website (Streamlit)

  1. (Recommended) create a virtual env

python -m venv .venv

# macOS/Linux:

source .venv/bin/activate

# Windows (PowerShell):

.venv\Scripts\Activate.ps1

  1. Install dependencies

pip install -r requirements.txt

  1. Launch the app

streamlit run app.py

  1. Open the local URL it prints (usually http://localhost:8501), upload an image, and explore:
  • Original + EXIF (editing software/date flags if present)
  • ELA heatmap for localized recompression
  • Noise residual for texture inconsistencies
  • Copy–move overlay showing repeated patterns with consistent displacement
  • JPEG Ghosts curve + variance map
  • Suspicion Score (0–1) with a clear verdict:
    • < 0.35 → Likely Unaltered (low suspicion)
    • 0.35–0.65 → Inconclusive (medium suspicion)
    • > 0.65 → Possibly Doctored (high suspicion)

You can also download a JSON report of all metrics from the app.

⚙️ Prefer the Command Line?

python cli.py –image /path/to/image.jpg –report out.json

Outputs a one-line verdict + a detailed JSON report.

✨ What’s going on under the hood?

  • EXIF: reads metadata (dates, camera, software strings); many PNGs/screenshots won’t have EXIF.
  • ELA: recompresses to JPEG and measures differences—strong localized artifacts can hint at edits.
  • Noise residual: high‑pass residual + patch heterogeneity to spot odd texture patterns.
  • Copy–move: ORB keypoints are matched within the image; a strong cluster of similar displacements can reveal cloning.
  • JPEG ghosts: recompress at several qualities, plot error curve, and compute a per‑pixel variance map across qualities—regions that behave differently can be suspicious.
  • Heuristic score: blends signals (ELA, copy–move, noise heterogeneity, JPEG variance) into a 0–1 score with a simple verdict.

🧭 Pro tips for better reads

  • Use the highest-resolution original you can get (avoid screenshots and re-uploads).
  • If the image is a PNG derived from a JPEG, ELA/JPEG tests weaken.
  • Heavy filters/compression can push false positives—always pair with context.
  • This is an educational tool; for high-stakes analysis, consult professional forensic workflows.

🌈 You’ve got this!

You now have a cheerful, no‑nonsense forensics starter kit—perfect for quick checks, demos, and learning.

If you want me to tailor the UI (brand colors, your logo, extra tests, or a hosted deployment guide), say the word and we’ll power it up! 🚀 

More posts