Reverse-Engineering SynthID is an advanced project focused on discovering, detecting, and removing Google's AI-generated watermark, SynthID, through sophisticated spectral analysis techniques. The SynthID watermark is an imperceptible pattern embedded within pixel values of images produced by Google's Gemini system. This project enables users to engage with the watermarking technology analytically and improve their understanding of watermark extraction methodologies through innovative signal processing.

Key Features

• Discovery of carrier structures: Identify the resolution-dependent frequency structure of the SynthID watermark.
• Detection capabilities: Develop a detector that boasts up to 90% accuracy in identifying watermarked images.
• Efficient bypass methods: Create multi-resolution spectral bypass techniques, achieving impressive results like a 75% drop in carrier energy and a phase coherence reduction of 91%.
• Multi-color consensus: Generalize the approach across various models and colors to ensure robust removal across different types of images.
• Result validation: Iterative testing has resulted in solutions that effectively bypass the SynthID detection systems.

Watermark Analysis

Several visual examples illustrate the watermark's impact. For instance, when amplifying high-frequency residuals on a plain white image generated by Gemini, the watermark becomes apparent:

Technical Overview

The project consists of a 7-stage attack pipeline, which includes:

1. VAE round-trip to project the image off the natural-image manifold.
2. Elastic deformation to disrupt the watermark's phase consistency across the image.
3. Global geometric operations like rotation and scaling to further obscure the watermark.
4. Squeeze-and-resize techniques to eliminate sub-pixel detail of the watermark.
5. Color and contrast adjustments to modify pixel statistics key to watermark detection.
6. FFT-phase subtraction using both universal and codebook-harvested carrier bins.
7. JPEG compression for additional disruption of the watermark signal.

Performance Metrics

In rigorous test scenarios, the project achieved near-lossless visual output while successfully bypassing detection systems. Key results include:

• A significant PSNR (Peak Signal-to-Noise Ratio) metric indicating high fidelity of output images.
• Successful detections on both gemini-3.1-flash-image-preview and nano-banana-pro-preview models across various image types.

Community Contribution

Contributions are welcome, especially pure black and white images generated by specific models to enhance detection and removal accuracy. Participating can help refine the watermark extraction process for future applications.

For further details on methodology and to download reference materials, visit the project page at PitchHut.

The ultimate aim of this research is to delve deeper into the realm of watermarking technology and enhance tools to study its robustness and vulnerabilities, contributing to ongoing discussions related to AI-generated content identification and security.