CSIRO’s Data61 latest research gives copyright owners and data producers a new superpower — the option to make their sensitive and sovereign data 𝘂𝗻𝗹𝗲𝗮𝗿𝗻𝗮𝗯𝗹𝗲 by AI. It’s been getting some attention, with us appearing on ABC TV, 7 News, and 2GB radio to explain the technology.

Unlike past “confuse the AI” tricks like data poisoning, which are basically a game of cat and mouse (and the AI is usually the cat), our approach comes with a 𝘁𝗵𝗲𝗼𝗿𝗲𝘁𝗶𝗰𝗮𝗹 𝗴𝘂𝗮𝗿𝗮𝗻𝘁𝗲𝗲. Add just the right amount of carefully calculated noise to an image, and no matter how clever or upgraded the AI becomes, it simply can’t learn the important features. That makes it 𝗳𝘂𝘁𝘂𝗿𝗲-𝗽𝗿𝗼𝗼𝗳.

This opens up a range of uses, from deepfake defence to protecting sensitive industrial images and intellectual property, while keeping image quality as high as possible.

We’re now looking for partners to help turn this into real-world, use-case-specific solutions. Winning a Distinguished Paper Award was nice, but seeing it start to make real-world impact is even better.

Paper: https://www.ndss-symposium.org/ndss-paper/provably-unlearnable-data-examples/