aloshdenny

reverse-SynthID

Security#AI#Reverse Engineering#Signal Processing#Watermarking#Python

131

// summary

This project reverse-engineers Google's SynthID AI watermarking system using spectral analysis and signal processing techniques without requiring access to proprietary encoders. It provides a robust detector with 90% accuracy and a V3 multi-resolution spectral bypass capable of significantly reducing watermark energy while maintaining high image quality. The system utilizes a multi-resolution spectral codebook to identify and surgically remove watermarks across various image resolutions.

// technical analysis

This project provides a framework for reverse-engineering Google's SynthID watermarking system through spectral analysis and signal processing techniques. By identifying that the watermark's carrier frequencies are resolution-dependent and share consistent phase templates, the project enables the detection and surgical removal of these invisible markers. The core technical innovation is the multi-resolution SpectralCodebook, which allows for precise frequency-bin-level subtraction while maintaining high image quality, effectively bypassing the watermark without relying on brute-force methods.

// key highlights

Achieves a 90% detection accuracy for SynthID watermarks using a robust, multi-scale extraction module.

Implements a V3 bypass that delivers 43+ dB PSNR, ensuring minimal visual degradation while removing the watermark.

Utilizes a multi-resolution SpectralCodebook to store and auto-select watermark fingerprints based on specific image dimensions.

Employs direct known-signal subtraction in the frequency domain, which is more effective than traditional noise injection or compression.

Provides a multi-pass iterative subtraction schedule to catch residual watermark energy across different strength levels.

Includes cross-validation tools that use pure black and white reference images to isolate and verify true carrier frequencies.

// use cases

Detecting SynthID watermarks in AI-generated images with 90% accuracy.

Removing SynthID watermarks using a multi-resolution spectral codebook bypass.

Analyzing the carrier frequency structure and phase consistency of AI-generated watermarks.

// getting started

To begin, clone the repository and install the dependencies using 'pip install -r requirements.txt'. You can then build a multi-resolution codebook from your own reference images or use the provided scripts to run the V3 bypass on target images. The project offers both a Python API for integration and a CLI for direct file processing.