Reverse-Engineering of SynthID Watermarking System

This repository explores the reverse engineering of Google's SynthID watermarking system, an invisible watermark embedded within images generated by Google Gemini. The project leverages signal processing and spectral analysis techniques without access to the proprietary encoder or decoder to achieve significant breakthroughs in watermark detection and removal.

Project Highlights

• Watermark Detection: Developed a detector capable of identifying SynthID watermarks with 90% accuracy.
• Spectral Analysis: Analyzed the resolution-dependent carrier frequency structure of the watermark for improved detection and removal strategies.
• Advanced Bypass Techniques: Introduced a multi-resolution spectral bypass (V3) achieving a 75% drop in carrier energy and a 91% drop in phase coherence, with over 43 dB PSNR across various image resolutions.

Contribution Solicitation

Contributions are welcome, particularly in the form of pure black (#000000) and pure white (#FFFFFF) images generated by Nano Banana Pro. These images are crucial for enhancing the multi-resolution extraction capabilities.

Unique Approach

The project stands out by employing a multi-resolution SpectralCodebook, which is a collection of watermark fingerprints for various resolutions, enabling precise frequency-bin-level removal tailored for any image size. This method contrasts traditional brute-force approaches such as JPEG compression or noise injection, providing a more targeted and effective solution.

Key Findings

1. Resolution-Dependent Watermarking: The watermarking system exhibits unique carrier frequency structures based on image resolution. A specific codebook profile is essential for accurate watermark removal.
2. Phase Consistency: The watermark's phase template remains consistent across images produced by the same model of Gemini, enhancing detection reliability.
3. Carrier Frequency Insights: Research into carrier frequency structures has yielded data critical to developing robust watermark detection strategies.

Technical Infrastructure

The repository includes several core modules for executing multi-resolution codebook creation and watermark detection:

• synthid_bypass.py: Implements the bypass techniques and manages the spectral codebooks for watermark removal.
• robust_extractor.py: Facilitates the detection of watermarked images.

The project aims to further the understanding of AI watermarking technologies, providing a resource for academic research and security analysis initiatives.