Has Google’s AI watermarking system been reverse-engineered?

A software developer, known as Aloshdenny, claims to have reverse-engineered Google DeepMind's SynthID AI watermarking system, making it possible to strip watermarks from generated images or insert them into other works. Aloshdenny has open-sourced their findings on GitHub, detailing a process that allegedly requires 200 Gemini-generated images, signal processing, and significant time, without using neural networks or proprietary access. The method involves generating numerous images, enhancing contrast, denoising, and averaging patterns to identify watermark frequencies, then attempting to remove them. However, Aloshdenny admits the process primarily confuses the decoder rather than completely removing the watermark, suggesting the system is designed to increase the cost of misuse rather than be unbreakable. Google, through spokesperson Myriam Khan, has denied these claims, stating that the tool cannot systematically remove SynthID watermarks and that the system is robust and effective for AI-generated content.

(Source：The Verge)