DGIST and Stanford University jointly develop one-shot federated learning AI technique combining privacy protection and efficiency

□ A research team led by Professor Sang-hyun Park of the Department of Robotics and Mechatronics Engineering at DGIST (President Kunwoo Lee), together with researchers from Stanford University in the United States, has developed a new one-shot federated learning artificial intelligence (AI) technique that enables efficient training of large-scale models without sharing personal information. This research outcome is extremely significant, as it demonstrates that privacy protection, learning efficiency, and model performance can be secured simultaneously in the field of medical image analysis.

□ Medical imaging data contain sensitive patient information, which restricts information sharing between hospitals and poses major challenges to developing AI models using large-scale datasets. Federated Learning (FL), proposed as a solution, enables collaborative training by sharing trained models instead of raw patient data. However, repeated model transmissions result in substantial time and cost burdens. One-shot FL has been studied as an alternative, but existing methods have continued to suffer from high computational costs and overfitting.

□ To address these limitations, Professor Park’s team proposed a method that adds structural noise to synthetic images and applies the mix-up technique to generate virtual intermediate samples. This approach increases training data diversity to reduce overfitting, while reusing synthetic images to eliminate unnecessary computation, thereby significantly improving computational efficiency.

□ The research team applied this technique to a range of medical imaging datasets, including radiographic images, pathological images, dermatoscopic images, and fundus images. The results showed that the proposed method achieved higher accuracy with fewer computations compared to existing one-shot FL approaches.

□ Professor Park commented, “This research is meaningful in showing that, even under realistic constraints such as privacy protection and communication limitations, it is possible to train broadly applicable models in the field of medical imaging. Moving forward, we will continue to advance this technique to develop AI models that encompass diverse patient populations while safeguarding privacy, thereby contributing to the establishment of accurate and highly reliable diagnostic support systems.”

□ This research was conducted as part of the DGIST General Research Program, the National Research Foundation of Korea’s Excellent Young Researcher Program, and the Ministry of Science and ICT’s AI Star Fellowship Support Program. The findings have been published in the October issue of Medical Image Analysis, one of the world’s leading international journals in the field of medical image analysis.