Study reveals best image disguising methods for medical AI privacy

A new study from researchers led by Jason Rojas systematically evaluates image disguising as a privacy-enhancing technology (PET) for cloud-based medical image analysis. With sensitive patient data often outsourced, disguising transforms images into unintelligible representations while preserving information for downstream learning. The team tested two representative methods—DisguisedNets and NeuraCrypt—on four datasets covering classification (e.g., disease detection) and dense semantic segmentation (e.g., organ delineation). Their unified framework assessed predictive utility, computational efficiency, and robustness against reconstruction attacks.

Results reveal a stark task dependency: disguising methods preserved utility well for medical image classification but caused substantial degradation in segmentation. Specifically, Randomized Multidimensional Transformation (RMT) emerged as the optimal approach, balancing performance and security. In contrast, AES-based disguising, while cryptographically strong, severely impacted model accuracy. Notably, regression-based reconstruction attacks that work on natural images proved considerably less successful on realistic medical images. These findings provide critical guidance for deploying PETs in confidential medical AI, highlighting that one-size-fits-all solutions fail and that task-specific evaluation is essential.