Skip to main content

Lecture 06 — Invisible ink

Watermarking synthetic media, and how quickly a robust-sounding signal turns out not to be

29 March 2027

Unlike a stylistic tell, a watermark is deliberate: a generator embeds a signal at creation time, precisely so that someone downstream can recover it. Google’s SynthID nudges the statistics of an image or a token sequence in a way that’s invisible to a viewer and detectable by the matching tool. It is, on paper, a much better answer than looking for accidental artefacts.

On paper. Researchers behind a disclosed attack called UnMarker reported knocking SynthID’s detection rate from roughly 100% down to about 21% — a 79% success rate at defeating the watermark — using nothing but a diffusion “purification” pass that regenerates the image through a generative model. They reported the result to Google under its Vulnerability Reward Program, and Google acknowledged it. Watermarks have to survive being screenshotted, recompressed, cropped, and re-uploaded through three platforms before anyone sees them again — and a watermark surviving all of that turns out to be a much harder property to guarantee than a watermark existing in the first place. We will implement a toy watermarking scheme, then spend the second half of the session trying to break each other’s.

Outline

Further reading

Teaching team