Reverse engineering Gemini's SynthID detection
github.com
April 9, 2026
7 min read
🔥🔥🔥🔥🔥
49/100
Summary
The reverse-SynthID project reverse-engineers Google's SynthID watermarking system, which is embedded in images generated by Google Gemini. It utilizes signal processing and spectral analysis to identify the watermark's resolution-dependent carrier frequency structure and has developed a detector that achieves 90% accuracy in watermark identification.
Key Takeaways
- The project reverse-engineers Google's SynthID watermarking system, enabling the detection and removal of the watermark embedded in images generated by Google Gemini.
- A detector has been built that identifies SynthID watermarks with 90% accuracy using spectral analysis.
- The V3 multi-resolution spectral bypass achieves a 75% carrier energy drop and 91% phase coherence drop, allowing for effective watermark removal across different image resolutions.
- The watermark's phase template is consistent across images from the same Gemini model, with the green channel carrying the strongest watermark signal.
Community Sentiment
Positives
- The initiative to improve multi-resolution watermark extraction shows a proactive approach to enhancing AI detection methods, which is crucial for maintaining trust in AI-generated content.
Concerns
- The ease of removing SynthID from AI-generated images raises serious concerns about the effectiveness of current detection methods, questioning their reliability in distinguishing real from AI-generated content.
- The low-quality nature of the AI-assisted research repository suggests a lack of rigor in testing against established detection systems, undermining confidence in the findings.