New 'Noise' Tech Protects Online Content from AI Learning
A groundbreaking new technique developed by Australian researchers offers a potential solution for individuals and organizations seeking to prevent unauthorized artificial intelligence systems from learning from their online content. The method, pioneered by CSIRO, Australia’s national science agency, in collaboration with the Cyber Security Cooperative Research Centre (CSCRC) and the University of Chicago, subtly alters image-based content—such as photos and artwork—making it unintelligible to AI models while remaining perfectly clear to the human eye.
This innovative approach holds significant promise for safeguarding digital assets and personal privacy in an era of pervasive AI. For content creators, it could act as a crucial barrier against intellectual property theft, while social media users might employ it to protect their personal images from being used to train AI systems or create sophisticated deepfakes. Imagine a social media platform automatically applying this protective layer to every uploaded photo, effectively curbing the rise of manipulated media. Similarly, defense organizations could shield highly sensitive satellite imagery or critical cyber threat data from being inadvertently absorbed into autonomous AI models.
What sets this technique apart is its underlying mathematical rigor. Unlike previous attempts that often relied on trial-and-error or assumptions about how AI models operate, this new method provides a verifiable guarantee that AI systems cannot learn from protected content beyond a specific, defined threshold. Dr. Derui Wang, a CSIRO scientist involved in the research, emphasized this distinction, stating that their approach offers a “powerful safeguard” for anyone uploading content online, asserting its resilience even against adaptive attacks or attempts by AI models to retrain themselves around the protection.
The scalability of the method is another key advantage. Dr. Wang highlighted its potential for automatic, widespread application by digital platforms. This means a social media site or other online service could seamlessly embed this protective layer into every image uploaded, offering a broad defense against data exploitation and unauthorized AI training. While currently focused on images, the research team has ambitious plans to extend the technique’s capabilities to other media types, including text, music, and video, signaling a broader vision for content protection across the digital landscape.
The method, though still theoretical, has shown promising results in controlled laboratory settings. Its underlying code has been made available on GitHub for academic use, fostering further research and development within the scientific community. The paper detailing this work, titled “Provably Unlearnable Data Examples,” garnered significant recognition, receiving the Distinguished Paper Award at the prestigious 2025 Network and Distributed System Security Symposium (NDSS), underscoring its potential impact. The research team is actively seeking partners across various sectors, including AI safety and ethics, defense, cybersecurity, and academia, to help translate this theoretical breakthrough into practical applications.