"Preventing AI from Recovering Deleted Memories"... UNIST Develops 'AI Memory Eraser' Technology [Reading Science]
Recognition Accuracy Suppressed to 0.24% Even After Relearning Attacks on Deleted Data
Potential Applications for Personal Data Protection
A domestic research team has developed an "AI memory eraser" technology that prevents artificial intelligence (AI) models from recovering deleted personal or copyrighted data, even if only a few photos are re-entered after the information was removed.
On July 21, UNIST (Ulsan National Institute of Science and Technology) announced that a team led by Professor Sung-Hwan Yoon of the Graduate School of Artificial Intelligence, together with Professor Saerom Park of the Department of Industrial Engineering, has developed a machine unlearning technology called "Spotter" that blocks even attacks attempting to relearn deleted information.
Comparison of the Impact of Prototypical Relearning Attack in Facial Recognition Models. Provided by the Research Team
View original imageMachine unlearning is a technology that selectively removes only the influence of specific data from a trained AI model. It is used to eliminate face information, personal data, or harmful content from face recognition AI when requested by individuals.
Blocking 'Relearning Attacks' That Restore Deleted Information
Existing unlearning technologies either suffered from "over-deletion," where even normal data similar to the deletion target was not recognized, or had the limitation that AI could relearn deleted information simply by re-entering a few target photos.
The Spotter technology developed by the research team selectively removes only the core criteria AI uses to recognize the deletion target, while maintaining the recognition performance of normal data similar to the deletion target. As a result, it can effectively block relearning attacks that attempt to recover deleted information.
In experiments using CIFAR-10, a representative AI training dataset containing images in 10 categories such as automobiles, cats, and airplanes, Spotter reduced the recognition accuracy for the deletion target to 0%. Even after a relearning attack using five deleted target photos, Spotter suppressed the recognition accuracy to just 0.24%. In contrast, existing unlearning technologies allowed the recognition accuracy for the deletion target to recover to as high as 99.98% under the same conditions. The recognition accuracy for other non-deleted categories was maintained at 99.96%.
The research team also confirmed similar performance on large-scale image datasets such as "Tiny ImageNet" and the face recognition dataset "CASIA-WebFace."
Research team photo. (From left) Professor Sung-Hwan Yoon, Professor Saerom Park, Researcher Seungbeom Ha. Provided by UNIST
View original imageThe research team expects this technology to be widely utilized in AI services where personal data protection is crucial, such as in the removal of personal information and harmful content from face recognition systems.
Ha Seungbeom, the first author of the study, stated, "Previously, we were unable to properly evaluate the decline in performance for normal data similar to the deletion target or the problem of deleted information being revived. Spotter is a technology that can complement the blind spots of current unlearning technologies."
Hot Picks Today
"Two Spend 500,000 Won in a Flash, Yet Wait Two Hours?" ... 'Dreamland' Shunned by Kids and Foreign Tourists
- The Era of Getting 200 Million Won After 30 Years Is Over... A Japanese Bank's "Perpetual Department Head" Who Just Clocked In Cries Out
- Spain Hit by Its Worst Tragedy... Lee Kangin, Moreno and Real Madrid Join Soccer World in Mourning
- They Said, “Make It Your Home for Life” - Then Suddenly Kicked 200 Passengers Off... What Happened to the Around-the-World Cruise?
- "130 Rice Bowls Sold Out in a Day"... 'Unfiltered Face' Adds 8 Million Won to Night Market Sales: China's Bungeunwoo
This research has been accepted for presentation at the International Conference on Machine Learning (ICML 2026), the world's most prestigious academic conference in the field of AI.
© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.