Paper: Adversarial Retain-Free Unlearning for Bearing Prognostics and Health Management
Authors: Hoki Kim (Chung-Ang University), Chaewon Yun (Chung-Ang University, undergraduate), Jiyong Lee (Chung-Ang University, undergraduate)
Venue: IEEE Transactions on Industrial Informatics (Top ~5% SCIE)
Original JoongAng Ilbo article
A technology has been proposed that enables industrial AI systems to selectively forget specific data. Even when data deletion requests arise, the existing model does not need to be retrained from scratch, as the problematic data can be removed. This is expected to simultaneously address privacy protection and data governance challenges in industrial settings.
Chung-Ang University announced that a paper by Prof. Hoki Kim’s research team from the Department of Industrial Security, titled ‘Adversarial Retain-Free Unlearning for Bearing Prognostics and Health Management,’ has been published in ‘IEEE Transactions on Industrial Informatics,’ a prestigious international journal in the field of industrial artificial intelligence. This journal is ranked in the top approximately 5% of SCIE journals and is recognized as a highly influential publication in the fields of industrial AI and smart manufacturing.
In industrial settings, AI-based systems are widely used for fault diagnosis of critical components such as bearings. However, when data deletion requests arise, the high cost of retraining due to large-scale data and complex model architectures makes practical implementation difficult in industrial environments. This issue has become an increasingly important technical challenge as privacy protection regulations and data governance requirements are strengthened.
To address this, the research team proposed a new framework called ARU (Adversarial Retain-Free Unlearning), which combines adversarial attacks and semantic-driven loss functions. This method generates ‘Surrogate Retain-like Samples’ from the pre-trained model and the data targeted for deletion, designed to maintain the model’s structural stability without using actual retained data.
Experimental results using public and industrial bearing datasets showed that the proposed ARU achieved higher forgetting performance and stable diagnostic accuracy simultaneously compared to existing machine unlearning techniques. This demonstrates that data deletion requirements can be efficiently handled even in actual industrial equipment diagnostic systems.
Prof. Hoki Kim of Chung-Ang University said, “As AI adoption expands across industries including smart factories, protecting data rights and ensuring reliability have become important challenges,” adding, “This research is meaningful in that it provides the first technical foundation for safely removing specific data in industrial AI environments.” He continued, “As AI applications expand, we plan to continue related research going forward.”
This research was supported by the National Research Foundation of Korea (NRF) and the Institute for Information and Communications Technology Promotion (IITP).