A Homomorphic Encryption Approach for Privacy-Preserving Deep Learning in Digital Health Care Service.

Applied deep learning technology in digital health care service is a potential way to tackle many issues that hospitals face, such as over health care requests, lack of doctors, and patient overload. But a conventional deep learning model needs to compute raw medical data for evaluating health information, which raises considerable concern about data privacy. This paper proposes an approach using homomorphic encryption to encrypt raw data to protect privacy while deep learning models can still perform computations over encrypted data. This approach can be applied to almost any digital health care service in which data providers want to ensure that no one can use their data without permission. We will focus on a particular use case (predict mental health based on phone usage routine) to represent the approach's applicability. Our encryption model's accuracy is similar to the non-encryption model's (only 0.01% difference) and has practical performance.