Researchers from CSIRO have developed a groundbreaking algorithm designed to prevent the misuse of images for creating deepfakes, particularly in the context of non-consensual sexualized content generated by AI.

This innovative algorithm, created in collaboration with the Cyber Security Cooperative Research Centre and the University of Chicago, subtly modifies images to make them unreadable to AI models while remaining indistinguishable to human viewers.

The algorithm can be implemented at scale, enabling social media platforms to automatically embed protective layers into uploaded images, which could significantly reduce the prevalence of deepfakes and intellectual property theft.

The research detailing this work, titled 'Provably Unlearnable Data Examples', was presented at the 2025 Network and Distributed System Security Symposium, where it received a distinguished paper award.

Although the algorithm is currently theoretical and has only been validated in laboratory settings, the team has made the code available on GitHub and is actively seeking partnerships to commercialize the technology.

Plans are underway to extend this protective technique to other media types, including text, music, and videos, broadening its potential impact.

In addition to combating deepfakes, this technique could protect artists' work from being used to train AI models, addressing ongoing debates about the use of copyrighted material in AI training.

CSIRO research scientist Dr. Derui (Derek) Wang highlighted that the method offers a mathematical guarantee against unauthorized machine learning models learning from the protected content.

This development comes at a crucial time, as Australian state governments, including those of Victoria, New South Wales, and South Australia, are moving to criminalize sexually explicit AI-generated content.