New technique helps prevent AI from learning from unauthorized data

Australian scientists have developed a new technique that could prevent Artificial Intelligence (AI) systems from learning illegally from images, artwork and other online content.

According to Australia's national science agency (CSIRO), the technique subtly alters the content of images to make them unreadable to AI models, while remaining unaltered to the human eye. This is a project developed by CSIRO, in collaboration with Australia's Cybersecurity Cooperative Research Centre (CSCRC) and the University of Chicago (USA).

The authors say the breakthrough could help artists, organizations, and social media users protect their work and personal data from being used to train AI systems or create deepfakes — incredibly realistic videos , images, or audio created by AI. For example, users could automatically apply a protective layer to images before posting, preventing AI from learning facial features to create deepfakes.

Similarly, defense organizations might protect sensitive satellite imagery or data about cyber threats.

According to Dr Derui Wang, a scientist at CSIRO, this method uses a strong mathematical foundation to ensure that AI models cannot learn from that content, or in other words, this technique makes the data "unlearnable" for AI to a level that protects privacy and copyright, while still retaining its usefulness to humans.

This protection remains in effect even if the AI tries to adapt or is retrained, he added.

Dr. Wang said the technique could be applied automatically at scale. He said a social media platform or site could embed this protective layer into all uploaded images. This could curb the proliferation of deepfakes, reduce intellectual property theft, and help users retain control of their content.

While the method is currently only applicable to images, the researchers plan to expand it to text, music , and video. The technology is still in the theoretical stage and has only been shown to work effectively in a lab setting.

The above scientific work, titled “Provably Unlearnable Data Examples”, was awarded the Outstanding Research Award at the 2025 Network and Distributed Systems Security Symposium (NDSS)./.

Source: https://www.vietnamplus.vn/ky-thuat-moi-giup-ngan-chan-ai-hoc-hoi-tu-du-lieu-khong-duoc-phep-post1055216.vnp