CSIRO, Australia’s national science agency, in collaboration with the Cyber Security Cooperative Research Centre and the University of Chicago, has developed an innovative algorithm designed to prevent Tracy from being used to create AI-generated deepfakes, particularly those of a sexually explicit nature.
This breakthrough, detailed in a distinguished paper presented at the 2025 Network and Distributed System Security Symposium, addresses growing concerns about non-consensual deepfake content, which has surged in popularity among various groups, including high school students.
The algorithm subtly modifies image pixels in a way that makes them unreadable to AI models while remaining visually unchanged to the human eye.
According to CSIRO research scientist Dr. Derui (Derek) Wang, this method provides a mathematical guarantee that protected images cannot be used to train AI models beyond a certain threshold, even if retraining is attempted.
Unlike previous approaches that rely on trial-and-error or assumptions about AI behavior, this technique offers a robust safeguard for individuals, artists, and organizations.
The potential applications of this technology are wide-ranging. It could help prevent the creation of deepfakes, protect artists’ intellectual property from being used to train AI, and shield sensitive imagery, such as defense-related satellite data, from being absorbed into AI models.
The algorithm could be integrated into platforms like social media, automatically applying a protective layer to uploaded images to curb deepfake proliferation and intellectual property theft.
While currently validated in a controlled lab environment, the algorithm’s code has been made available on GitHub for academic use.
CSIRO aims to collaborate with researchers and industry partners to bring this technology to commercial applications.
The team also plans to extend the method to protect text, music, and videos, potentially transforming the field of AI content protection.
The content has been paraphrased and summarized under fair use principles for educational and informational purposes, with full attribution to the original source.
