Reverse Engineering Google Gemini's SynthID Watermark: 90% Detection Accuracy Without Access to Encoder — 167 Points

Hacker News·medium signal

A GitHub project demonstrates that Google's SynthID watermarking system can be reverse-engineered using only signal processing and spectral analysis, achieving 90% watermark detection accuracy and 75% carrier energy removal. The research found SynthID's phase template is identical across all images from the same Gemini model, with the green channel carrying the strongest signal. This has significant implications for AI-generated content detection — if watermarks can be reliably detected and removed without access to proprietary tools, the entire watermarking approach to AI content provenance may be fundamentally flawed.