Privacy-Preserving Medical Image Segmentation via Hybrid Trusted Execution Environment
Event Type
Special Session (Research Track)
Virtual Programs
Hosted in Virtual Platform
Machine Learning/AI
DescriptionThe strict privacy requirements placed on medical records by various government regulations has become one of the major obstacles in deploying machine learning algorithms in medicine. Recently, it is reported that the-state-of-the-art secure protocol is able to segment a three-dimensional cardiovascular CT image without revealing any sensitive information related to the medical image involved in the computation. However, it takes more than 3,000 seconds to segment a single image, which is much longer than the time needed by an experienced radiologist and makes it clinically impractical. In this talk, we make use of the trusted execution environment (TEE) built upon a software/hardware co-exploration approach to implement a privacy-preserving medical image segmentation protocol. We show that the proposed approach, while preserving privacy and accuracy, only needs less than 60 seconds to segment an image (over 50x faster than the state-of-the-art), for the first time achieving higher efficiency than human experts.