🕵️‍♂️ Reverse SynthID: AI Watermark Forensics Lab

A deep learning and computer vision research tool designed to extract, visualize, and analyze hidden imperceptible watermarks (specifically targeting Google's SynthID) embedded in AI-generated images.

This project provides both a Python/Flask backend utilizing scientific computing libraries (SciPy, PyWavelets) and a Client-Side (OpenCV.js) implementation for rapid browser-based analysis.

🔬 The Science Behind the Tool

SynthID and similar invisible watermarking technologies embed data within the high-frequency components of an image. This tool reverse-engineers the visual traces of these watermarks using a three-step forensic pipeline:

1. Noise Extraction (High-Pass Filtering): - Python: Utilizes Discrete Wavelet Transform (db4 wavelets via PyWavelets) to separate the low-frequency structural image data from the high-frequency residual noise.
   • JavaScript: Uses a Difference of Gaussians (DoG) approximation to isolate the noise layer.
2. Residual Amplification: The extracted noise is heavily amplified (e.g., 50x) to make the imperceptible pixel variations visible to the human eye, often revealing grid-like or wave patterns.
3. Frequency Domain Analysis (FFT): A Fast Fourier Transform (FFT) converts the spatial noise map into a frequency spectrum. By shifting the DC component to the center and applying a logarithmic scale, the tool reveals "star patterns" or bright dots.
4. Pattern Matching: The tool overlays known "secret" carrier frequencies (derived from research) onto the FFT spectrum. If the bright spots in the frequency domain align exactly with these targeted coordinates, it indicates a high probability of SynthID presence.

📁 Project Structure

The repository contains two distinct implementations:

1. Python Flask Application (/app.py, /analysis_engine.py)

A robust server-side application for deep analysis.

• app.py: The Flask web server handling uploads and routing.
• analysis_engine.py: The core scientific pipeline performing wavelet denoising and FFT generation.
• requirements.txt: Python dependencies.
• templates/ & static/: Frontend UI and storage for generated analysis maps.

2. Client-Side JS Visualizer (/SynthID-Visualizer/)

A lightweight, browser-only tool leveraging OpenCV.js.

• index.html: The UI for the client-side tool.
• js/visualizer.js: Handles image processing, FFT, and canvas rendering entirely in the browser.
• assets/carriers.json: The coordinate map of known SynthID carrier frequencies.

🚀 Installation and Usage

Running the Python/Flask App

1. Clone the repository and navigate to the project directory.
2. Install the required dependencies:

   pip install -r requirements.txt

(Requires: Flask, numpy, opencv-python, scipy, PyWavelets, matplotlib)

3. Start the Flask server:

   python app.py

4. Open your browser and navigate to http://127.0.0.1:5000/. Upload an image to see the noise residual and frequency fingerprint.

Running the Client-Side Visualizer

1. Navigate to the SynthID-Visualizer directory.
2. Serve the directory using any local web server to avoid CORS issues with the local JSON file. For example:

   python -m http.server 8000

3. Open http://localhost:8000/index.html in your browser.

📊 Interpreting the Results

When an image is analyzed, you will see two primary outputs:

• Extracted Noise: A static-like image. Look for repeating diagonal or grid-like textures hidden within the noise.
• Frequency Fingerprint: A heatmap of the frequency domain.
  • Look for the Green Circles (the known carrier frequencies).
  • If bright white/yellow dots appear perfectly centered inside these green circles, the image exhibits the signature of a Gemini/SynthID watermark.

⚠️ Disclaimer

This is an educational research tool designed to demonstrate computer vision and signal processing techniques. It relies on approximated client-side and open-source methods to visualize frequency patterns and is not an official Google SynthID decoding tool.