Privacy-Preserving Deep Learning With Learnable Image Encryption on Medical Images

The need for cloud servers for training deep neural network (DNN) models is increasing as more complex architecture designs of DNN models are developed. Nevertheless, cloud servers are considered semi-honest. With great attention to the privacy issues of medical diagnoses using a DNN, previous studies have proposed the idea of learnable image encryption. Though some methods have been presented to partially attack previous encryption schemes, there is still some space for improvement. We proposed a learnable image encryption scheme that is an enhanced version of previous methods and can be used to train a great DNN model and simultaneously keep the privacy of training images. We conducted an experiment on medical datasets from open sources and the result demonstrates the effectiveness of our proposed method in performance and privacy-preserving.