Published 17 hours ago • loading... • Updated 1 hour ago
Google Figures Out How to Watermark AI-Designed Proteins
The method could help DNA-screening software identify AI-designed proteins without changing their function, researchers said.
On Wednesday, Google DeepMind researchers published a paper in Nature1 detailing SynthIDBio, a watermarking system that embeds hidden signatures into AI-designed proteins to reveal their machine-generated origins.
AI-Designed tools like AlphaFold and RFdiffusion can create novel proteins that current biosecurity screens often miss because they lack counterparts in existing databases. This gap complicates threat identification.
Computer scientist Pushmeet Kohli and his team at DeepMind 'stress-tested the approach on a number of challenging problems' to ensure the watermark remains effective without compromising protein function.
DNA synthesis providers could use a secret detection key to identify watermarked proteins, aiding biosecurity screening. Tessa Alexanian, a biosecurity researcher, said 'Any information you can give us to make sense of these orders is good.'
While watermarking supports innovation and security, it is not foolproof. Someone can run a watermarked protein through another design tool and generate a new sequence that obscures its synthetic origins.
A new, open-source tool that embeds digital watermarks into artificially produced DNA and proteins, enhancing global biosecurity without altering their functions.