Survey: Privacy-Preserving in Deep Learning based on Homomorphic Encryption

Abstract

When deep learning techniques succeed, the amount of data accessible for training grows rapidly, the main successor is the collecting of user data at scale in huge enterprises. Because users' data is sensitive, and the manner of preserving this information (pictures and audio recordings) indefinitely, data collecting raises privacy concerns. The terms privacy and confidentiality are related to avoiding sharing this information, and deep learning cannot be accrued on a larger scale in the amount of data. There are some challenges in machine learning algorithms when needing to access data for the training process. There several technologies in deep learning for privacy-preserving have been evolving to assign the issues, including the multi-lateral computation secrecy and the symmetric encryption in the term of the neural network. This survey deals with the deep learning techniques concerning the privacy issue mainly related to input data and the ability of interesting directions in these learning processes. Finally, as a side contribution, we analyze and introduce some variations to the bootstrapping technique of deep learning. That offers an improved parameter in efficiency at the cost of increasing privacy.