Towards Privacy-Preserving Deep Learning: Opportunities and Challenges
Sheraz Ali, Muhammad Maaz Irfan, Abubakar Bomai, Chuan Zhao · 2020
During the past decade, deep learning has achieved excellent results in many classic machine learning problems, such as face recognition, spam detection, and financial prediction, etc. The success of deep learning mainly benefits from training on a large amount of data gathered from different sources. However, the training data may include highly sensitive information that may lead to serious threats to personal privacy. In this paper, we review privacy threats in deep learning and present effective privacy-preserving techniques in the literature against these threats. We give a detailed comparison of different techniques and summarize the performance of the existing solutions. Furthermore, we discuss some open problems and challenges in this research area. In the end, we conclude this paper and point out possible research directions.