Privacy-preserving deep learning in medical informatics: applications, challenges, and solutions

Deep Learning (DL) has already shown tremendous potential in designing intelligent clinical support systems in biomedicine. Data privacy plays a significant role while training and testing DL models, especially for sensitive data. Privacy-Preserving Deep Learning (PPDL) applications in Healthcare are rapidly growing as medical informatics deals with sensitive data. This work reviews the recent advances in PPDL techniques in Healthcare. It first analyzes the need of PPDL in healthcare informatics using a threat model and then discusses privacy-preserving computation techniques for secure data processing and evaluation. Next, it focuses on DL applications over Healthcare in three categories: (i) PPDL in the private cloud, (ii) PPDL in the public cloud, and (iii) privacy based on modifications in DL architectures. Next, we examine data privacy at different stages of DL deployment in Healthcare, including input, model, training, and output. We also provide a summary of the evaluation outcomes of the solutions reviewed. Additionally, we highlight the unique challenges in PPDL for Healthcare and offer suggestions for future research directions.